Automated Stratigraphic Correlation

FURTHER TITLES IN THIS SERIES

1. A.J. Boucot
EVOLUTION AND EXTINCTION RATE CONTROLS

2. W.A. Berggren and J.A. van Couvering

THE LATE NEOGENE - BIOSTRATIGRAPHY, GEOCHRONOLOGY AND
PALEOCLIMATOLOGY OF THE LAST 15 MILLION YEARS IN MARINE AND
CONTINENTAL SEQUENCES

3. L.J. Salop
PRECAMBRIAN OF THE NORTHERN HEMISPHERE

4. J.L. Wray
CALCAREOUS ALGAE

5. A. Hallam (Editor)
PATTERNS OF EVOLUTION, AS ILLUSTRATED BY THE FOSSIL RECORD

6. F.M. Swain (Editor)

STRATIGRAPHIC MICROPALEONTOLOGY OF ATLANTIC BASIN AND
BORDERLANDS

7. W.C. Mahaney (Editor)

QUATERNARY DATING METHODS

8. D. Jan6ssy
PLEISTOCENE VERTEBRATE FAUNAS OF HUNGARY

9. Ch. Pomerol and I. Premoli-Silva (Editors)

TERMINAL EOCENE EVENTS

10. J.C. Briggs

BIOGEOGRAPHY AND PLATE TECTONICS

11. T. Hanai, N. lkeya and K. lshizaki (Editors)

EVOLUTIONARY BIOLOGY OF OSTRACODA. ITS FUNDAMENTALS AND
APPLICATIONS

12. V.A. Zubakov and 1.1. Borzenkova

GLOBAL PALAEOCLIMATE OF THE LATE CENOZOIC
Developments in Palaeontology and Stratigraphy, 13

Automated Stratigraphic
Correlation

El? Agterberg
Mathematical Applications in Geology Section, GeologicalSurvey of Canada,
601 Booth Street, Ottawa, Ont., K 1A OE8, Canada

ELSEVIER
Amsterdam - New York - Oxford -Tokyo 1990
ELSEVIER SCIENCE PUBLISHERS B.V.
Sara Burgerhartstraat 25
P.O. Box 21 1, 1000 AE Amsterdam, The Netherlands

Distributors for the United Stares and Canada:

ELSEVIER SCIENCE PUBLISHING COMPANY INC.

655, Avenue of the Americas
New York, NY 10010, U S A .

ISBN 0-444-88253-7

0 Elsevier Science Publishers B.V., 1990

All rights reserved. No part of this publication may be reproduced,.stored in a retrieval system or
transmitted in any form or by any means, electronic, mechanical, photocopying, recording or
otherwise, without the prior written permission of the publisher, Elsevier Science Publishers B.V./
Physical Sciences & Engineering Division, P.O. Box 330, 1000 AH Amsterdam, The Netherlands.

Special regulations for readers in the USA -This publication has been registered with the Copyright
Clearance Center Inc. (CCC), Salem, Massachusetts. Information can be obtained from the CCC
about conditions under which photocopies of parts of this publication may be made in the USA. All
other copyright questions, including photocopying outside of the USA, should be referred to the
copyright owner, Elsevier Science Publishers B.V., unless otherwise specified.

No responsibility is assumed by the Publisher for any injury and/or damage to persons or property
as a matter of products liability, negligence or otherwise, or from any use or operation of any meth-
ods, products, instructions or ideas contained in the material herein.

This book is printed on acid-free paper.

Printed in The Netherlands

 V

FOREWORD
Geological correlation of strata plays a key role in sedimentary basin
analysis. Such correlation, particularly when scaled in linear time,
requires that a series of unique points for non-recurrent events like
occurrences of fossils must first be determined, common to t h e
sedimentary record as observed a t different sites. An important
contention of geological correlation is that once such events, probably
grouped in biozones, have been properly determined and defined, these
units can indeed be used for correlation. This statement, which might
seem to be trivial, is made here because existing stratigraphic codes show
how to construct stratigraphic units but they do not define how to correlate
them. The actual correlation generally takes place in the subjective
domain of regional experts on a particular basin o r time period.
Procedures for correlation or stratigraphic equivalence depend on
subjective evaluation of the unique relation of each individual site record
to the derived and accepted standard. It follows that correlation as
practiced in geology cannot be readily verified without a detailed, and
probably exhaustive review of all the underlying facts. Traditionally
there is no method of formulating the uncertainty in fixation of individual
records t o the standard. Hence biostratigraphy often is more considered
an art rather than a science. The problem of using subjective judgement
only is not so much that it leads to right or wrong stratigraphy, but that a
single solution is proposed. It should be attempted to establish reasonable
criteria for successful correlation by providing insight into the actual
uncertainty in correlation, either in millions of years or in depth in meters.

This book is an important review on 25 years of progress in computer-

based stratigraphic correlation of fossil data. The best methods should
combine sound mathematical logic with sound stratigraphic reasoning,
and allow the user to retain full control over input and results. The author
of this study is at the forefront of research and development i n
quantitative stratigraphy, particularly with respect t o methods that apply
to fossil distributions as frequently found in exploration wells in frontier
basins. The ten chapters systematically explore the foundations and
objective applications of quantitative biostratigraphy. This will bring us a
step closer to a more automated procedure of correlation, applicable in a
wide range of sedimentary basin analyses.

F.M. Gradstein, Chairman,

Committee on Quantitative Stratigraphy,
Dartmouth, Nova Scotia, January 1990
This Page Intentionally Left Blank
 VZI

PREFACE
The purpose of this book is to provide an introduction t o recent
developments in automated stratigraphic correlation using computer
programs for ranking and scaling of stratigraphic events. It is intended for
advanced geology students, research workers and teachers with a
background in stratigraphy and a n interest in using computer-based
techniques for problem-solving. The mathematical background provided
is sufficient to justify the methods that are used but the equations are
relatively few and concentrated in specific sections (mainly in Chapters 3,
6 and 8) and may be skipped by readers who are not mathematically
inclined. Occasionally, use is made of elementary statistical techniques
(t-test, chi-squared test or analysis of variance) on which additional
explanations can be found in one of the numerous excellent introductory
textbooks on probability and statistics in existence.
After data inventory for a region or time period, the stratigrapher
first proceeds to establish a regional zonation which later can be used for
correlation. Age calibration is a requirement for constructing this
zonation as well as for the process of stratigraphic correlation. The
computer can play a n integral r81e in these procedures. In this book, the
emphasis is on worked-out examples of application of ranking, scaling and
correlation of stratigraphic events using relatively small datasets, for
illustration of the intermediate steps made within the computer between
input and output. It should be clear t o the reader that automated
stratigraphic correlation is not a simple automatic process such a s
alphabetic sorting. The stratigrapher has to integrate vast amounts of
information which cannot possibly be stored in large databanks. Every
piece of evidence or link between different pieces of evidence or hypotheses
has its own sources of uncertainty associated with it. Using a computer for
problem-solving may violate uncertainties that cannot be quantified.
Computer input, therefore, always should be evaluated critically by expert
stratigraphers and paleontologists.
In total there are ten chapters. The purpose of the first two chapters
is to introduce the probabilistic method for automated stratigraphic
correlation and t o discuss principles of quantitative stratigraphy.
Applications of mathematical statistics and computer science not
specifically dealing with ranking and scaling but of interest t o
stratigraphers and paleontologists are presented in Chapter 3. Coding and
file management of stratigraphic information (Chapter 4) provides the
VlII
input required for ranking and scaling of biostratigraphic events by means
of the RASC method treated in the next two chapters. A number of topics
including rank correlation, precision of the scaled optimum sequence,
normality testing and t h e modified RASC method a r e presented
separately (in Chapters 7 and 8) as extensions and refinements of the
RASC method. The chapter on event-depth curves a n d multi-well
comparison (Chapter 9) contains examples of regional applications with
automated correlation between stratigraphic sections. Finally, in Chapter
10, much of the material on methods presented in earlier chapters is
summarized in a general description of t h e micro-RASC system of
computer programs for ranking, scaling and regional correlation of
stratigraphic events.
I a m indebted to many individuals and organizations for support.
Foremost among these is Felix Gradstein of the Atlantic Geoscience
Centre of the Geological Survey of Canada who started me thinking about
automated biostratigraphic correlation in 1978. From 1979 to 1986, I had
t h e privilege of being t h e Leader of Project 148 ( Q u a n t i t a t i v e
Stratigraphic Correlation Techniques) of the International Geological
Correlation Programme co-sponsored by Unesco and the International
Union of Geological Sciences. This project and later the Committee on
Quantitative S t r a t i g r a p h y of t h e I n t e r n a t i o n a l Commission on
Stratigraphy provided the framework for regular discussions with most
colleagues active in method development for quantitative stratigraphy. I
have used suggestions of m a n y of t h e s e colleagues, especially
P.O. Baumgartner (UniversitB de Lausanne, Switzerland), G.F. Bonham-
Carter (Geological Survey of Canada, Ottawa), J.C. Brower (Syracuse
University, Syracuse, New York, U.S.A.), J.M. Cubitt (Poroperm, Chester,
U.K.), E. Davaud (Universitb de Genkve, Switzerland), P.H. Doeven
(Petro-Canada, Calgary, Canada), C.W. Drooger (University of Utrecht,
the Netherlands), L. Edwards (U.S.G.S., Reston, Virginia, -U.S.A.),
C.M. Griffiths (University of Trondheim, Norway), J. Guex (Universitb de
Lausanne, Switzerland), C.W. Harper, Jr. (University of Oklahoma,
Norman, U.S.A.), W.W. Hay (University of Colorado, Boulder, Colorado,
U.S.A.), I. Lerche (University of South Carolina, Columbia, S.C., U.S.A.),
D.F. Merriam (Wichita State University, Wichita, Kansas, U.S.A.),
M. Rube1 (Academy of Sciences, Estonian SSR, Tallinn, U.S.S.R.),
W. Schwarzacher (Queen's University, Belfast, U.K.), B. S t a m (Shell
Syria, Damascus), J.E. Van Hinte (Free University, Amsterdam, t h e
Netherlands) and M. Williamson (Shell Canada, Calgary, Canada).
 IX

Thanks are due to these individuals for their critical remarks during
development of the ranking and scaling techniques to be discussed. I am
grateful for assistance by computer programmers at the Geological Survey
of Canada especially to Ning Lew, Louis Nel and Jacqueline Oliver, and t o
Dan Byron, Marc D’Iorio, and Kazim Nazli as my students at the Ottawa-
Carleton Geoscience Centre.
For this book I have made extensive use of material in publications
authored or co-authored by me during the past 10 years. On eight
occasions, I was one of the lecturers of the one-week Quantitative
Stratigraphy Short Course given under the auspices of IGCP Project 148
and the Committee on Quantitative Stratigraphy in Canada (2 X 1, Brazil,
China, Holland, India, U.K. and U.S.A. Mostly attended by stratigraphers
and quantitative geoscientists from oil companies, this course provided a
stimulating environment for jointly exploring and testing ideas on how to
use computers intelligently. Those familiar with the earlier work will find
many extensions of the RASC method made during the past three years
especially in the fields of coding the original stratigraphic information,
comparison with other methods and statistical evaluation. For example, it
was well known that ranges on average range charts constructed by means
of RASC tend to be shorter than those resulting from most other methods.
The new modified RASC method yields range charts with wider ranges
connecting entries to exits for taxa in those stratigraphic sections where
these taxa were observed at their lowest and highest positions relative t o
all other taxa considered.
The Geological Survey of Canada has allowed me t o work on this book
project which involved extensive support including drafting and
photography. The project would not have been possible without the
invaluable help in word-processing received from Janet Gilliland, Shirley
Kostiew, Guylaine Leger and Diane Winsor. Martin Tanke of Elsevier has
provided guidance and encouragement. Last but not least I thank my wife
Codien for her help and understanding.
F.P. Agterberg,
Ottawa, January 1990
This Page Intentionally Left Blank
 XI

CONTENTS

Foreword ...................................................... V
Preface ...................................................... VII

CHAPTER1. PROBABILISTIC M E T H O D F O R A U T O M A T E D
STRATIGRAPHIC CORRELATION
1.1 Introduction ............................................. 1
1.2 IGCPProject 148 ........................................ 2
1.3 Quantitative biostratigraphy ............................. 5
1.4 Quantitative chronostratigraphy ......................... 11
1.5 Quantitative lithostratigraphy ........................... 14
1.6 Recent developments in stratigraphy ..................... 15

CHAPTER 2 . PRINCIPLES OF QUANTITATIVE STRATIGRAPHY

2.1 Introduction ............................................ 19
2.2 Zones in biostratigraphy ................................. 20
2.3 Quantitative versus qualitative stratigraphy .............. 26
2.4 Local versus regional ranges of taxa ...................... 30
2.5 Estimation of the highest and lowest occurrences of taxa .... 31
2.6 The frequency distributions of highest and lowest
occurrences of taxa ...................................... 37

CHAPTER 3. APPLICATIONS O F MATHEMATICAL STATISTICS

AND COMPUTER SCIENCE TO ZONATION.
CORRELATION AND AGE INTERPOLATION
3.1 Introduction ............................................ 47
3.2 Binomial test for randomness ............................ 48
3.3 Binomial distribution model for microfossil abundance data . 49
3.4 Multiple pairwise comparison ............................ 60
3.5 Applications of graph theory ............................. 61
3.6 Use of cubic smoothing splines for removing
“noise” from microfossil abundance data .................. 67
3.7 Biostratigraphic correlation between Tojeira 1and 2 sections
in central Portugal using E . mosquensis abundance data .... 70
3.8 Multivariate methods ................................... 73
3.9 Research on time-scales ................................. 76
3.10 Computer simulation experiments on estimation of
the age of chronostratigraphic boundaries ................. 85
XI1
3.11 Smoothing of time-scales with the aid of cubic
spline functions ......................................... 92
3.12 Statistical significance of ages ............................ 98

CHAPTER4 . CODING AND F I L E MANAGEMENT OF

STRATIGRAPHIC INFORMATION
4.1 Introduction ........................................... 103
4.2 Five basic types of files ................................. 103
4.3 Hay example as derived from the Sullivan database:
Lower Tertiary nannoplankton in California ............. 108
4.4 Partial DAT file for the Hay example .................... 112
4.5 DAT files constructed by Guex and Davaud ............... 116
4.6 Gradstein-Thomas database: Cenozoic Foraminifera
in Canadian Atlantic Margin wells ...................... 118
4.7 Characteristic features of Gradstein-Thomas database ..... 125
4.8 Frequency of occurrence of taxa of Cenozoic Foraminifera
along the northwestern Atlantic margin ................. 129
4.9 Artificial datasets based on random numbers ............. 132

CHAPTER 5 . RANKING OF BIOSTRATIGRAPHIC EVENTS

5.1 Introduction ........................................... 141
5.2 Hay’s original method .................................. 142
5.3 Algorithmic version of Hay’s original method ............. 145
5.4 Uncertainty ranges for events in the optimum sequence ... 152
5.5 Other ranking algorithms .............................. 154
5.6 Conservative ranking methods .......................... 165
5.7 Three-event cycles ..................................... 170
5.8 Higher-order cycles and pseudo-cycles ................... 174
5.9 The influence of coeval events ........................... 175

CHAPTER 6. SCALING OF BIOSTRATIGRAPHIC EVENTS

6.1 Introduction ........................................... 179
6.2 Scaling versus ranking ................................. 183
6.3 Statistical model for scaling of stratigraphic events ........ 186
6.4 Artificial example ..................................... 201
6.5 Computer simulation experiments ....................... 204
6.6 Normality test ......................................... 215
6.7 Marker horizon option of the RASC method ............... 219
6.8 Unique event option of RASC program ................... 221
6.9 Binomial and trinomial models for scaling ................ 223
 XI11
6.10 Application of Glenn and David’s trinomial model ......... 227
6.11 Comparison of observed and estimated probabilities ....... 236

CHAPTER 7. RANK CORRELATION AND PRECISION OF SCALED

OPTIMUM SEQUENCE
7.1 Introduction ........................................... 239
7.2 Rank correlation coefficients ............................ 239
7.3 RASC step model ...................................... 242
7.4 Presorting and ranking by Harper ....................... 246
7.5 Precision of the scaled optimum sequence ................ 250

CHAPTER 8. NORMALITY TESTING AND THE MODIFIED RASC

METHOD
8.1 Introduction ........................................... 259
8.2 Autocorrelation of the second-order differences ........... 260
8.3 Unitary Associations and RASC methods applied to
Drobne’s alveolinids .................................... 268
8.4 Application of RASC and normality test to Palmer’s
database for the Riley Formation in central Texas ......... 276
8.5 Modified RASC method ................................. 280
8.6 Application of modified KASC to the Gradstein-Thomas
database .............................................. 284
8.7 Frequency distributions of stratigraphic events ........... 287
8.8 Application of modified RASC to Drobne’s alveolinids ..... 295
8.9 Comparison of range charts for Palmer’s database ......... 305

CHAPTER9. EVENT-DEPTH CURVES AND MULTI-WELL

COMPARISON
9.1 Introduction ........................................... 311
9.2 Principles of correlation and scaling in time and
comparison to composite standard method ................ 312
9.3 Generalized description of the CASC method ............. 320
9.4 Statistical selection of optimum spline-curves ............. 338
9.5 Cross-validation method ................................ 339
9.6 Jackknife method ...................................... 342
9.7 Computer simulation experiment for event-depth
spline fitting with error analysis ........................ 347
9.8 Regional application of RASC and CASC ................. 351
9.9 Application of RASC and CASC t o Hibernia Oilfield ....... 358
9.10 Application of CASC t o Palmer’s database ................ 366
x IV
9.11 Benthic foraminiferal zonation, central North Sea . . . . . . . . . 371
9.12 Integration of foraminiferal and dinoflagellate datasets,
Labrador Shelf-Grand Banks . . . . .. . . . . . . . . . . . . . . . . . . . . . . 382

CHAPTER 10.COMPUTER PROGRAMS FOR RANKING, SCALING

AND REGIONAL CORRELATION OF STRATIGRAPHIC
EVENTS
10.1 Introduction . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . 389
10.2 Summary of contents of the 12 modules of micro-RASC . . . . 391
10.3 List of decisions to be made by user of the RASC
computer programs . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . 396
10.4 Brief history of the development of RASC and CASC . . . . . . . 404

REFERENCES ....................................... ........ 409

INDEX . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
 1

CHAPTER 1

PROBABILISTIC METHOD FOR AUTOMATED

STRATIGRAPHIC CORRELATION

1.1 Introduction

From 1976 to 1986 about 150 scientists in 25 countries collaborated

under the auspices of the International Geological Correlation Programme
in Project 148: Evaluation and Development of Q u a n t i t a t i v e
Stratigraphic Correlation Techniques. More recently similar work has
been performed within the context of the Committee for Quantitative
Stratigraphy of the International Commission on Stratigraphy. Although
individual paleontologists and stratigraphers had used quantitative
methods before, the collaboration in IGCP-148 led to new mathematical
methods of stratigraphic correlation, mainly in biostratigraphy but also in
chronostratigraphy and lithostratigraphy. These methods are reviewed in
this book with emphasis on those developed by the author and his
colleagues in Canada.

Sequencing methods deal with the relative order of stratigraphic

events such as the highest occurrences of fossil taxa as observed in many
sections. Intervals between successive events in an ordered sequence can
be estimated (scaling) and the results expressed in linear time if a
subgroup of the stratigraphic events can be dated. Such methods have
been used extensively, e.g. t o construct biozonations for Jurassic and
younger sediments along the NW Atlantic margin (Gradstein et al., 1985)
and, recently, t o develop a new deep water benthic foraminifera1 zonation
for the Cenozoic strata of the Central and Viking Grabens, North Sea
(Gradstein et al., 1988; Agterberg and Gradstein, 1988). Several regional
hiatuses of 2 t o 5million years (Ma) in duration, stand out and match
changes in sea level. The same methods have been employed for
automated isochron contouring with error bars in depth o r time units in
Cenozoic and Cretaceous basins, off eastern Canada. Such information
may be used for automated basin history analysis.
2

Time-successive assemblages of fossils also can be established by

using multivariate methods on co-occurrences of events or with Guex’s
(1987) method of Unitary Associations in conjunction with graph theory on
the overlap of stratigraphic ranges. Other methods for stratigraphic
correlation to be reviewed in this book include Shaw’s (1964) composite
standard method and various uses of cubic spline functions for smoothing
and interpolation. Attractions of quantitative stratigraphy are the use of
rigorous methodology which highlights many properties of the data, the
ability to handle large and complex data bases in an objective manner, and
statistical evaluation of the uncertainty in the results. Generally, little
conceptual orientation is required in order t o use these methods and
thereby gain more information from a particular dataset.

1.2 IGCP Project 148

The IGCP Project “Evaluation and Development of Quantitative

Stratigraphic Correlation Techniques” was initiated in 1976 for the
purpose of developing computer-based mathematical theory and analysis
of geological information which can be applied t o obtain automated
correlation techniques in stratigraphy. These techniques are especially
important in analysis of hydrocarbons and coal bearing basins. The
project was terminated in 1986 and final results were described in
Agterberg and Gradstein (1988). The rapid growth of data in stratigraphy
has led to an increased demand for quantification of the data for machine-
handling and graphic display. Quantitative stratigraphy is useful in this
because it helps t o organize the data in novel ways. Specific problems can
be solved by establishing regional standards of ordered stratigraphic
events and performing correlations on the basis of these standards
preferably with estimates of uncertainty. Comprehensive descriptions and
computer programmes have been prepared for different techniques which
were applied t o the same datasets in order to evaluate their respective
advantages and drawbacks. The purpose of these evaluations is to select
those techniques which are relatively simple and easily understood,
achieve maximum resolution also in comparison with traditional methods
of stratigraphic correlation, and can be implemented on computers of
different types including microcomputers.

Studies in the fields of biostratigraphy, lithostratigraphy (especially

well logs) and sedimentology make successful use of the quantitative
 3

modelling approach. Stztistical and other numerical techniques can be

used for erection of biozonations, correlation of zones and events,
classification and matching of lithofacies in well logs or sections,
lithofacies pattern recognition, and modelling of geological processes
relative t o the numerical time scale. The IGCP-148 participants were
conducting research mainly in the fields of biostratigraphy and
lithostratigraphy. Special attention was given t o the performance of
computer-based quantitative techniques in comparison with the results
obtained by conventional qualitative stratigraphic correlation methods.

During the first years of existence (1976 to 1981), the emphasis

within IGCP-148 was on method development. The statistical problems
encountered when attempting t o describe quantitative methods of
stratigraphic correlation in a cohesive manner are far more complex and
difficult to solve than one might expect. Some of the studies made under
the auspices of IGCP-148 would not have been possible without recent
advances in the theory of mathematical statistics, especially graph theory
for order relationships between stratigraphic events or co-occurrences of
fossil species, and spline-curve fitting theory for age-depth relationships
with error analysis.

Later the primary activity in IGCP-148 shifted from method

development to application, for solving specific stratigraphic problems
using large data bases for regions in North America, Europe and India.
Deep Sea Drilling Project data sets in the Atlantic and Pacific Oceans were
also analyzed. Except for subprojects on the Silurian in the Baltic region
and the Cambrian in Texas, the participants have been working mostly on
Cenozoic, Cretaceous and Jurassic stratigraphy.

Research on the following major problems was mostly completed:

Creation and definition of a mathematical theory of stratigraphic

relationships.

Establishment of standards and codes for the biostratigraphic,

lithological and environmental information attainable from well logs,
cores, and surface sections.

Development of a mathematical theory for stratigraphic correlation.

Development of practical methods of biostratigraphic correlation

concentrating on quantification of assemblage zones, sequencing
4

methods, set theoretical approaches, morphometric chronoclines and

multivariate methodology.

Development of practical methods of correlation concentrating on

methods of spectral analysis (frequency domain), methods of
stretching and zonation (time domain), methods of stratigraphic
interpolation and multivariate statistical analysis.

Over 200 publications emanating from IGCP-148, including computer

programs, have been listed in Geological Correlation and the IGCP
Catalogues. This includes collections of papers in books and special issues
of scientific journals (Cubitt, Editor, 1978; Gill and Merriam, Editors,
1979; Cubitt and Reyment, Editors, 1982; Agterberg, Editor, 1984;
Gradstein et al., 1985; Agterberg and Rao, Editors, 1988; Oleynikov and
Rubel, Editors, 1989). After 1986, the international co-operation achieved
was continued under the auspices of the Committee on Quantitative
stratigraphy of the International Commission on Stratigraphy which
recently has provided an indexed list of 637 publications on quantitative
biostratigraphy (Thomas et al., 1988). For other recent papers see
Agterberg and Bonham-Carter ( 1 9 9 0 , P a r t 111: Q u a n t i t a t i v e
Stratigraphy).

A comprehensive review of quantitative biostratigraphy for the

period 1830-1980 already had been published by Brower (1981). Tipper
(1988) reviewed 400 articles in the general field of quantitative
stratigraphic correlation providing a n annotated bibliography. Both
Brower (1981) and Tipper (1988) noted t h a t the development of
mathematical techniques has tended t o outstrip their acceptance by
practicing stratigraphers. It is true that sophisticated techniques not only
require more mathematical background from the user but, if not used
knowledgeably, could lead to unrealistic or erroneous results more readily
than simple methods. On the other hand, techniques that are easy to
understand may be too simplistic for application in the real world. The
best methods should provide new insights by combining mathematical
logic with sound stratigraphic reasoning and allowing the user to retain
full control over input and output.

In the International Stratigraphic Guide of the Subcommission on

Stratigraphic Classification of the International Commission on
Stratigraphy (Hedberg, Editor, 1976) a clear distinction is made between
 5

(1) Lithostratigraphy in which strata are organized into mappable units

based on their lithologic character;

(2) Biostratigraphy with correlative units based on fossil content of

strata; and

(3) Chronostratigraphy with superimposed units based on the relative

age relations of the strata.

In this book, as in IGCP Project 148, emphasis is on biostratigraphy, a field

in which relatively few quantitative methods were available 12 years ago.

In order to explore the relation between qualitative and numerical

methods, this book is started with a review of principles and definitions in
stratigraphy in this chapter and the next one, emphasizing the biosphere
record.

1.3 Quantitative biostratigraphy

Numerical methods in biostratigraphy make use of the quantified

fossil record in sedimentary rock sections for precise recording and
correlation of extinct biological events in space and time. They can be
grouped into six basic categories:

Sampling and delineation of environments with fossils that occur in

patches (instead of displaying random spatial distributions);

Automated microfossil recognition;

Analysis of evolutionary sequences;

Measurement of the attributes of index fossils;

Determination of the most likely (scaled) sequence of biostratigraphic

events as recorded in different stratigraphic sections; and

Analysis of assemblage zones and concurrent range zones.

Emphasis in this book is on subjects (11, (5) and (6). This includes the
construction of range charts depicting periods of existence for different
fossil taxa in comparison with one another.
6
There are few basic studies that shed light on the actual distribution
of fossils in rocks from a statistical point of view. For a review and
applications t o modern benthic Foraminifera and Late Cretaceous
molluscs, see Buzas et al. (1982). The geological factors affecting the
chance of event detection generally remain unknown and cannot be
modelled prior to extensive sampling and stratigraphic analysis itself. On
the other hand, it is widely known from repeated observations that for
many groups of organisms, the majority of taxa is found a t relatively few
sampling sites and with few specimens. Figure 1.1 shows the cumulative
number of highest or lowest occurrences of taxa in well o r outcrop sections
in different areas of a large number of taxa of Mesozoic radiolarians,
Cenozoic dinoflagellates, Cenozoic Foraminifera and Cretaceous
nannofossils. The radiolarian and nannofossil data use lowest and highest
occurrences; the dinoflagellates and foraminifers highest occurrences only.

The graphs of Figure 1.1 show that the number of lowest or highest
occurrences of taxa found in at least 1 , 2 , 3 , ..., n sites, decreases steadily.
In other words, the majority of species (events) occur at few sites and few
species (events) are ubiquitous. It is noted that the sections used for the
examples vary in density and spacing and the shapes of the curves in
Figure 1.1 are influenced by methods of sampling. In Figure 1.1,
dinoflagellate events are most localized and nannofossils least. The use of
first and last occurrences increases traceability of taxa as shown for the
radiolarians and nannofossils. Obviously, quantitative stratigraphic
methods may want t o cull the data so as t o avoid use of species for which
the number of events is limited and enhances “noise”. Thresholds in, for
example, ranking and scaling (RASC) are set such that no use is made of
events that occur in less than h, sections; h, is set by the user. Rare events
of value for age calibration can be re-introduced later, during final
analysis.

Several computer-based methods are available for determining the

most likely sequence of biostratigraphic events recorded in different
stratigraphic sections and for the construction of quantitative range
charts. The resulting zonations can be of either the average or
conservative types. In general, average zonations will underestimate the
position of the highest occurrence of a range zone a t a given place while
they overestimate its base. On the other hand, the concept of an average is
tied to that of a probability distribution. This allows bases and tops t o be
fitted with confidence limits (see later). Conservative zonations are
produced by sequencing methods designed to give the stratigraphically
 7

NUMBER OF WELL SECTIONS

Fig. 1 . 1 Cumulative frequency distributions of stratigraphic first and last occurrences of microfossils in
Mesozoic and Cenozoic strata: 1 = number of dinoflagellates occui ring in 2, 3, ... wells; data for 249 last
occurrences of Cenozoic dinoflagellates in 19 wells, northwestern Atlantic margin; 2 = data for 119 first
and last occurrences of late Cretaceous nannofossils in 10 wells, northwestern Atlantic margin; 3 = data
for 220 first and last occurrences of Mesozoic radiolarians at 76 sites, Mediterranean and Atlantic
realms; 4 = data for 116 last occurrences of Mesozoic foraminifers in 16 wells, northwestern Atlantic
margin; 5 = data for 147 last occurrences of Cenozoic foraminifers in 29 wells, central North Sea (from
Agterberg and Gradstein, 1988).

highest possible estimate of t h e top of a range zone a n d t h e

stratigraphically lowest estimate of the base of a range zone. Their
drawback is that they are sedsitive to anomalous situations arising when,
locally, fossils were moved upwards or downwards in a stratigraphic
section due to mixing of sediments later in geological time or because of
contamination. When a fossil was poorly preserved, misidentification may
also be a reason that its range of occurrence in a section is under- or
overestimated. Assemblage zones, concurrent range zones and other types
of zones are easily derived from dissecting the sequence of all events.
Assemblage zones can also be determined by means of multivariate
statistical methods such as cluster analysis. In the latter methods, the
order of successive events in time is not used but zonations are obtained
from co-occurrencesof different species in the samples.

A new approach (Unitary Associations method; see later) developed

during the past 12 years by J. Guex and E. Davaud in Switzerland uses
graph theory t o establish the order relationships of events formed by
overlap of stratigraphic ranges. The final associations are mathematically
successive assemblages of fossil ranges which are equivalent t o the Oppel
zones of traditional biostratigraphy (Guex, 1987). Baumgartner (1984)
employed the Unitary Associations method to propose a comprehensive
Tethyan radiolarian zonation with 14 zones in 43 Middle Jurassic - Early
Cretaceous sections. All zones are defined and identified in the sections.
Several zones would not have been detected without the quantitative
method employed for this study mainly because of patchiness of the fossil
record.

Special properties of the paleontological record form the basis of bio-

stratigraphy. These properties include first appearance datum (entry),
range, peak occurrence, and last appearance datum (exit) of fossil taxa.
Paleontological correlation for geological studies depends on comparing
similar fossil occurrences in or between regions by means of a
paleontological zonation. The observed order of paleontological events is
generally different from place to place. In correlating wells drilled for oil,
occurrences of the same event in different wells normally are connected by
straight lines in stratigraphic profiles or fence diagrams. If there is a
reversal in order for two events in two wells, these lines will cross. The
cross-over frequency for pairs of events, therefore, provides a measure of
inconsistency.

During the late 1950s and early 1960s’ Shaw (1964) had developed a
simple semi-objective method (Composite Standard method) of the
conservative type for dealing with inconsistencies. First and last
appearances of paleontological events in two sections are plotted against
each other. Next a line is fitted by using the method of least squares and
used for combining the two sections (line of correlation). The updated
positions of first or last appearances are those that are respectively lower
or higher in either of the two sections. A new section is plotted against the
combination of the first few sections. The procedure of adding other
sections is repeated until the “composite standard” is obtained that reflects
the maximum ranges of taxa. Shaw’s (1964) methodology was to a large
extent based on original work by earlier quantitative paleontologists,
notably Brinkmann (1929) who introduced basic concepts of statistical
biostratigraphy .

Shaw’s approach continues to be widely used. There is similarity

between it and the methods advocated in this book. The RASC approach
first gives a composite standard and lines of correlation are constructed
later. Computer-based variants of Shaw’s method include those developed
by Edwards (1984; 1989) and Gradstein and Fearon (1990). Edwards’
method is computer-based in that the stratigrapher combines sections and
subjectively fits lines while displaying intermediate results on the screen
 9

of a computer terminal. The method of Gradstein and Fearon is

microcomputer-based and employs De Boor’s (1978) cubic splines for
curve-fitting. In both methods intermediate results can be modified until
a satisfactory composite standard is obtained a t the end of a session.
So-called probabilistic methods which produce average ranges view
biostratigraphic sequences as random deviations from a true solution. The
solution faces four sources of uncertainty:

(1) The uncertainty due t o the fact that the optimum, or “true”, sequence
of fossil events has not been established. Under the influence of
Hay’s(1972) paper, ranking of events in time t o arrive a t their
stratigraphic order i s often referred t o a s “Probabilistic
Stratigraphy”. Binomial theory was used to evaluate superpositional
relations between events for statistical significance. However, as
Agterberg and Nel(1982a,b) have pointed out, there are no simple
models t o rank stratigraphic events according t o a numerical
probability. The problem is that order in time should be based both on
direct and on indirect estimates. For example, in Hay’s binomial
theory the fact that event A occurs above B in several sections ranks
the same as that A in some sections occurs above events C, D, E, F and
G, and that in some other sections C, D, E, F and G occur above B.
Both situations lead to the conclusion that A occurs above B, although
there is no simple way t o express this in terms of numerical
probability and more advanced mathematical methods for multiple
comparison have to be used.

(2) The uncertainty due t o the fact that the intervals between fossil
events along a relative time scale are not known (spacing or scaling
problem). In conventional biostratigraphy extensive use is made of
distances in time between events or (non) overlap of ranges t o produce
assemblage zones. In the simple, graphical technique of the
composite standard as developed by Shaw (1964), distance between
two or more successive events is a function of the relative dispersion
of each event in the sections considered; first occurrence levels are
minimized and last occurrence levels are maximized, but no direct
standard errors are available for the composite positions.

(3) The uncertainty due t o the fact that the geographic distribution of an
event is not known. Drooger (1974) refers to this as traceability. As
pointed out earlier, few taxa are ubiquitous and most species are rare.
10

Consequently, recovery is strongly affected by the vagaries of lateral

change in facies. Nevertheless, given enough sampling points and
counts, interpolations may be used to predict the potential presence of
each species.

(4) The error in the determination of biostratigraphic events at the scale

of a well, or outcrop section. This is basically a sampling error which
calls for an understanding and mathematical expression of errors in
field and laboratory techniques.

In order t o arrive at an optimum zonation and to attach confidence limits

t o correlations, considerable quantitative insight into these four sources of
uncertainty is required.

For the purpose of coping with numerous inconsistencies in a data-

base, containing many benthonic Foraminifera in wells along t h e
Canadian Atlantic margin (see Section 4.7),a computer program for the
ranking and scaling of events (RASC program) was developed by the
author in collaboration with F.M. Gradstein and co-workers in Canada
which produces three types of biostratigraphical answers:

The optimum (or average) sequence of stratigraphic events along a

relative time scale.

The clustering in relative time, of these events, based on the cross-

over frequencies of the events, weighted for t h e number of
occurrences, using the optimum sequence of (a)as input. This results
in a scaled optimum sequence with variable distance interval
between each pair of successive events along the RASC scale.

The stratigraphic and statistical normality (or comparison of order

relationships) of the events in individual sections compared with the
scaled optimum sequence.

In large-scale applications, the RASC computer program h a s

produced range charts and assemblage zonations which superseded
micropaleontological resolution-previously available. For example,
D’Iorio (1986) used this method for integration of large Cenozoic
foraminifera1 and dinoflagellate datasets from wells drilled on the Grand
Banks and Labrador Shelf, northwestern Atlantic Margin. In comparison
with optimum sequences for Foraminifera and dinoflagellates taken
separately, an increase in stratigraphic resolution of the regional biozones
 11

and a minor reordering of successive events resulted from this process of

integration (see Section 9.12). Although a dataset for a single fossil group
is enlarged when microfossils from other groups are added, the gain in
statistical precision because of larger sample sizes may be counteracted by
the introduction of new sources of bias related t o differences in
environmental control and completeness of information, between the
different fossil groups.

1.4 Quantitative chronostratigraphy

An approach i n which b i o s t r a t i g r a p h y , paleoecology,

lithostratigraphy, and geochronology are combined with one another is
called burial history (cf. Stam et al., 1987) or geohistory analysis (Van
Hinte, 1978; also see Lerche, 1990). It deals with subsidence and
sedimentation in time. Data from wells or sections are organized linearly
with the rates of subsidence, sedimentation and thermal maturation of
organic matter, expressed in years, thousands of years, o r larger time
units. Special emphasis is placed on a method for decompaction of
subsurface sedimentary units, using sonic logs or porosity data.

The prerequisite of this approach is a good calibration of fossil

zonations with respect t o the geochronologic scale. The determination of
trends is the primary objective and individual errors in calibration are less
important. This is because the trends can be generalized and used for
extrapolation, whereas errors in calibration produce localized “noise”
which should be eliminated if possible.

Information on rates of sedimentation, change in paleo-waterdepth,

unconformities, and other factors can be integrated in time with sediment
thickness data and paleo-waterdepth plots (cf. Doveton, 1986).
Refinements include corrections for compaction and loading which provide
information on seafloor or basement subsidence, evaporite movements,
undercompaction phenomena and exact timing of important changes in
geological history. The linear time perspective significantly clarifies
geological history and therefore exploration geology. This is primarily so
because it allows “dynamic” reconstruction of sedimentary basin history,
e.g. the time of maturation and migration of hydrocarbons in a region may
be postulated in linear time.
12

“Explorationists” also can establish a numeric chronostratigraphy for

well sections and calculate estimates for the extent in time of the missing
section a t unconformities (cf. Van H i n t e , 1978; Mohan, 1985).
Consequently, a new kind of cross-section can be constructed that shows
isochrons imaging chronostratigraphic depositional patterns just like the
seismic record does. As their geochronologic resolution normally will be
higher than that of seismic sections, isochron cross-sections are most
useful in the calibration and the interpretation of the seismic record.

As a follow-up t o the RASC (ranking and scaling) program, a

computer-based method of quantitative correlation was proposed, which
uses a numerical geologic time scale resulting from RASC. The computer
program is called CASC (Correlation And Scaling in time). Both
mainframe and microcomputer versions of CASC have been developed.
The mainframe version (Agterberg et al., 1985) provides two types of
displays. Initially, a n event-depth curve is constructed for each
stratigraphic section or well considered. Later the results for different
sections are correlated.

Figure 1.2 shows a CASC multi-well comparison for five offshore

wells on the Labrador Shelf. Briefly, the method runs as follows. A
separate set of biostratigraphic events (exits of microfossils only) was
observed in each well. By using the RASC computer program, a scaled
optimum sequence was obtained for a group of 2 1 wells. The RASC
distances of 54 events each occurring in 7 or more wells were transformed
into ages in millions of years using a subgroup of 23 Cenozoic
foraminifera1 events for which literature-based ages were available. This
allowed the construction of event-depth curves for individual wells. A
probable age can be computed for any point along the depth-scale of a well,
together with an error bar expressing the uncertainty of this estimate.

Three types of error bars are shown in Figure 1.2. A local error bar is
estimated separately for each individual well. It is two standard
deviations wide and has the probable isochron location a t its center. Use is
made of the assumption that the rate of sedimentation is linear in the
vicinity of each isochron computed. Consideration of nonlinear
sedimentation rates results in the asymmetrical modified local error bar of
Figure 1.2B. Like the local error bar a global error bar (Fig. l.2C) is
symmetric but it is based on estimates of uncertainty in age which are
 13

computed from the uncertainty in distance of the 54 foraminifera1 events

in the scaled optimum sequence based on all (21) wells.
In a large-scale application, Williamson (1987) used the Ranking and
Scaling method t o erect eleven biozones for the Hibernia oil field region,
Grand Banks, Canada (also see Chapter 9). Using the CASC method for a
regional time-scale interpretation of the zonation and isochron correlation,
Williamson proposed a subsurface correlation framework t h a t t o a
considerable extent matches the results of subsurface seismic sequence
analysis and provides chronostratigraphic correlation. He pointed out
that these computer programs put many of the concepts and philosophies
that have been used for many years by biostratigraphers on a statistical
basis, and as such, prospective users of the techniques would require little

Fig. 1.2 Example of CASC multi-well comparison with three types of error bar. The probable positions of
the time-lines were obtained from event-depth curves fitted to the biostratigraphic information of
individual wells. For further explanation see text.
14

conceptual orientation in order t o use these methods and thereby gain

more information from a particular data set.

1.5 Quantitative lithostratigraphy

Lithostratigraphic correlation can be defined a s the correct
identification of lithological boundaries in different locations. When the
correlated points are connected, they reproduce the shape of the rock body
(lithosome). This type of correlation is not probabilistic and, in the
stratigraphic sense, it is not even measurable. By establishing
quantitative methods, a probability measure of whether a proposed
correlation is right or wrong may be found. The similarity between two
sections is a measurable quantity. If two portions in the sections are
identical, this can be called a match and the number of matches is used as
a measure of the similarity. An example of a simple matching technique
for estimating the similarity between two successions of lithologies is to
divide the number of matches by the total number of comparisons made.
This technique called “cross-association” is explained in detail by Davis
(1986, pp. 234-239). Elaborating on these concepts, Vrbik (1985) obtained
statistical properties of the number of runs of matches between two
random stratigraphic sections. Olea (1988) has developed an interactive
computer system for lithostratigraphic correlation of wireline logs.

A fundamental prerequisite for such quantitative approach is the

meaningful numerical coding of lithologies. In addition, most quantitative
modelling studies require interpolation between equal intervals. This can
be accomplished by linear interpolation between irregularly spaced points
along sections or by using more sophisticated tools such as the cubic spline
function. Smoothing factors in spline interpolation can be determined by
interactively using a computer terminal, or by employing statistical
methods such as cross-validation (see Section 9.5).

Because of differences in the rate of sedimentation, stretching or

shrinking of sections is normally required before lithostratigraphic
correlation is possible (cf. Mann and Dowell, 1978; Shaw, 1978; Kwon and
Rudman,1979; Kemp,1982). An example of a new technique is the
slotting method for pairwise comparison of sections (cf. Gordon, 1982).
Suppose that two sections with observed lithological parameters, Al, A2,
..., An and B1, B2, ..., Bn are t o be slotted. One series, e.g. Al, A2, B1, A3,
B2, A4, A5, ..., can be created in which the successive data points show a
 15

minimum of dissimilarity. This method works best with continuous

lithological variables as obtained in well logging (Gordon and Reyment,
1979). Clark (1989) has developed a randomization test for comparison of
ordered sequences obtained by slotting or other matching techniques.
In addition t o differences in rate of sedimentation, hiatuses can
present a problem in lithostratigraphic correlation. Smith and
Waterman (1980) introduced a stratigraphic correlation algorithm
designed to deal with the gap problem. This technique was originally used
in studies of evolution of genetic sequences in molecular biology
(Waterman et al., 1976). Their approach is also closely related to “time-
warping” in speech recognition (Sankoff and Kruskal, Editors, 1983). An
essential property of these methods is the ability t o include gaps in
correlations. A single stratigraphic unit can be made a gap (not matched)
and several adjacent units can be treated as a single gap. The single-gap
method was programmed by Howell(1983). In its most general form
(Waterman and Raymond, 19871, one o r several adjacent strata in a
column can be matched with one or several strata in a second column and
deletions within one of these multiple matches also are possible. The latter
new algorithms include a method of minimum distance and a method of
maximum similarity. Within this context, a similarity algorithm is given
to locate and correlate the best matching segments or intervals from each
lithostratigraphic column considered.

1.6 Recent developments in stratigraphy

Radiometric methods provide estimates of age in millions of years.
However, any radiometric method is subject to a measurement error which
is usually much greater than the uncertainties associated with the
relative ordering of events using methods of stratigraphic correlation (e.g.
biostratigraphic or magnetopolarity methods). Relatively imprecise
isotope determinations can be combined to produce more precise estimates
of the age of stage and chronozone boundaries (cf. Section 3.9).
Recently, the International Commission on Stratigraphy has
published a global stratigraphic chart with geochronometric and
magnetostratigraphic calibration (Cowie and Bassett, 1989) incorporating
information of numerous subcommissions, working groups and
committees. A considerable amount of uncertainty remains associated
with some stage boundaries mainly because different radiometric methods
16

SEA LEVEL rel. Present (rn) 6 l80 PDB

-90-1 OOm
m

.
-200 -100 0 100 200 300 3.0 2.0 1.0 0.0 -1.0 -2.0
I I I

-
<
I I I I I I
0 , I ---__-__ Plio-Pleistocene 1
104
20 - 20 -
1 Miocene

A
0
Oligocene
2
Y
30-@’

gc 40-
Eocene
50 -

60
I
i
I
Crelaceour

70
’O’

Fig. 1.3 Comparison of the magnitudes of sea level events of the Tertiary as inferred by Vail et al. (1977)
from seismic stratigraphy, and the composite benthic 6 1 8 0 record according to Miller and Fairbanks
(1985). The encircled numbers refer to particular rises and falls examined by Williams et al. (1988). Also
see Table 1.1.

may yield results that are significantly different. For example, Odin
(1982) estimated the age of the Jurassic-Cretaceous boundary at 130 f 3
Ma but Harland et al. (1982) obtained 144 f 5 Ma. These 95 percent
confidence intervals do not overlap indicating unresolved problems of
methodology. This subject will be discussed in more detail in Section 3.12.
Menning (1989) has provided a synopsis of 30 complete and partial
geochronological time scales for the Phanerozoic published over a 70-year
period t o 1986. It is remarkable how close the most recent time scales are
to the first scale of Barrel1 (1917). For example, Barrell’s estimate of the
Jurassic-Cretaceous boundary was 135 M a which is identical to the age
estimate for this boundary in the above-mentioned 1989 global
stratigraphic chart. On the other hand, many geologists prefer the 144 Ma
estimate of Harland et al. (1982) and Kent and Gradstein (1985) for the
age of the Jurassic-Cretaceous boundary (cf. Section 3.12).
Seismic stratigraphy and isotope chronostratigraphy (Williams et al.,
1988) are providing new tools for the stratigrapher. For example, Figure
1.3 is a comparison of the magnitude of particular sea level events of the
Tertiary as inferred from seismic stratigraphy (Vail et al., 1977) and the
 17

composite benthic 6l80 record (Miller and Fairbanks, 1985). The two
patterns exhibit a similar long-term trend. Table 1.1 (after Williams et
al., 1988) compares magnitudes of 8 Tertiary sea level events (rises or
falls) based on the two methods. These are 3rd order events. In almost all
instances, the inferred sea-level change using sequence boundary patterns
yielded larger estimated changes than the 6 l 8 0 signal. The overall
agreement is not good a t this level of detail but both these types of
methodology are new and subject t o continuous improvement. For a recent
review of this topic and other approaches of chemical stratigraphy t o time-
scale resolution, see Williams (1990).

Quantitative dynamic stratigraphy (cf. Cross, Editor, 1990) is the

application of mathematical procedures to the analysis of geodynamic,
stratigraphic, sedimentogic and hydraulic attributes of sedimentary
basins. These are viewed as features produced by the interactions of
dynamic processes operating on physical configurations of the Earth at
specific times and places. A typical model of this type may represent
currents of water in sedimentary basins that alternately erode, transport
and deposit sediments. These processes can be represented by means of
differential equations t h a t are solved repeatedly with numerical
parameters which control their rate. Philosophies and strategies of model
building in this field are discussed by Lerche (1990).

TABLE 1.1
Comparison of the magnitude of particular sea level rises and falls based on seismically
defined unconformities with the 8180 record (after Williams et al., 1988, Table 11,
p. 112).

Event Type Timing(Ma) Agreement Seismic(m) 818O(m)

fall 15.5-6.6 poor -300 <50
fa11 24 good < 50 < 50
rise 30-15.5 poor > 300 < 100
fall 30 poor > 400 < 50
fall 52-37 poor < 100 -250
fall 40 good -100 -100
fall 59 poor < 150 < 50
fall 62.5 poor -200 < 50
This Page Intentionally Left Blank
 19

CHAPTER 2
PRINCIPLES OF QUANTITATIVE STRATIGRAPHY

2.1 Introduction
The original meaning of stratigraphy is “description of layers” and
like most earth science disciplines it is essentially a natural philosophy.
This implies t h a t stratigraphy is rooted in a body of organized,
historically-accumulated observations, governed by a series of widely
accepted principles and rules. The t w o physical principles of this
philosophy are:

1) geological time is irreversible because it is directed along the arrow of

time; and

2) sedimentary layers are laid down sequentially, one after another and
become younger upwards if left undisturbed (law of Steno; cf. Nowlan,
1986).

Over the last 200 or more years the science of stratigraphy has developed
into several major categories of effort and knowledge.

Lithostratigraphy is concerned with the classification, description

and lateral tracing or matching of rock units, characterized mainly by
their physical properties like sediment-type, degree of fossilization and
alteration, texture, and color. Modern techniques for classification also
make use of properties like seismic velocity (seismostratigraphy), or
emission and propagation of a host of physical signals in boreholes (log
analysis). The principal problem that besets classification and tracing or
matching (whether automated or not) is that lithological characteristics
are non-unique and repeat themselves in geological time. As a result,
there is a fundamental difference between the quantitative treatment of
single sections and quantitative approaches to lithostratigraphic tracing
based on multiple comparison of sections. Since the principal unit of
lithostratigraphy is the formation, which is a so-called mappable unit of
distinctive lithology, it is more appropriate to use tracing as a proof of
original continuity of strata, rather than correlation, which should be
reconstructed from biostratigraphy or magnetostratigraphy. Correlation
20

requires that a series of unique points for non-recurrent events must first
be determined, common t o the stratigraphic record as observed a t different
sites. An excellent introduction to this field of study is by Schwarzacher
(1985a,b).

The properties of the paleontological or fossil record form the basis of

biostratigraphy, which generally is called upon t o determine the unique
points of correlation, mentioned earlier. In the stratigraphic record the
paleontologist recognizes fossil taxa and from the continuous change of
taxa through time stratigraphic events are reconstructed. A taxon is
defined as a stable unit consisting of all individuals (fossils) considered to
be morphologically sufficiently alike to be given the same (Linnean)name.
For stratigraphic purposes, a taxon (species, or unit of different rank) is
recognized by a qualified paleontologist, whether based on single
specimens or “populations”. Commonly, categories intermediate between
such taxa are not used.

Biostratigraphic events are defined by the presence of a taxon in its

time context,-as derived from its position in a rock sequence. For
stratigraphic purposes relatively few events per taxon are considered only,
such as the first occurrence (appearance, entry), the last occurrence
(disappearance, exit), and possibly the most common or peak occurrence
between an entry and an exit. These events are the result of the evolution
of life on Earth. They differ from physical events in that they are unique,
non-recurrent, and that their order is irreversible. As a result, the three-
fold division of geological time into (1)prior to, (2) during, and (3) after the
existence of a taxon, is not ambiguous and provides a basic tool for
stratigraphic correlation. It is implied that each taxon was potentially
present at all points in time between its entry and exit. Absences within
its range are either environmental or preservational. This principle for
constructing ranges also was discussed by Cheetham and Deboo (1963).
Subsequent authors (cf. Brower, 1981; Tipper, 1988) referred t o it as the
“range-through” method.

2.2 Zones in biostratigraphy

The principal unit of “measurement” in biostratigraphy is the zone. A
zone is a body of strata commonly characterized by the presence of certain
fossil taxa. The most common types of zones are (after Hedberg, ed., 1976):
(1) assemblage zone ----- a group of strata characterized by a distinctive
 21

interval

I 11
zone

concurrent
rangezone

- 'I1 range zone

assemblage
zone B

assemblage
zone A

multi-taxon
concurrent
range zone

Fig. 2.1 Types of zones commonly used for biostratigraphic correlation (simplified from Hedberg, Editor,
1976). See text for further explanation.

assemblage of fossil taxa; (2) range zone ----- a group of s t r a t a

corresponding t o the stratigraphic range of a selected taxon in a fossil
assemblage; (3) concurrent range zone ----- the overlapping part of the
range zones of two or more selected taxa. The use of two or more taxa
whose range zones overlap reinforces correlation; (4)phylo-zone ----- a
body of strata containing a segment of a morphological-evolutionary
lineage for a taxon, defined between the predecessor and the successor.
The taxon is part of a lineage with morphologically well defined
increments assumably in stratigraphic order; and (5) interual zone ----- the
stratigraphic interval between two successive biostratigraphic events. In
general, zones based on drill cutting samples are interval zones.

Several types of zones are schematically represented in Figure 2.1.

Assemblage zones, multi-taxon concurrent range zones and Oppel zones
are based on many taxa. The taxa in assemblage zones may have lived
together or were accumulated together under similar conditions.
22

Assemblages may recur in a stratigraphic sequence and then can be useful

as indicators of environments. They may represent a given geological age,
although they are not controlled by the end points of ranges of taxa. In
general, evolutionary changes have been sufficient t o make assemblages
of one age distinctive from those of another age. Multi-taxon concurrent
range zones and Oppel zones both are based on the endpoints of ranges of
taxa. According to Hedberg (Editor, 1976), the concept of the Oppel Zone
largely embodies the concept of the concurrent-range zone but relaxes its
strict interpretation sufficiently to allow supplementary use of
biostratigraphic criteria other than range-concurrence that are believed to
be useful for demonstrating time equivalence. Thus the Oppel zone is
more subjective, more loosely defined and more easily applied than the
concurrent range zone.
The techniques to be described in this book are automated so that
large databases can be treated by computer-based statistical techniques
using stratigraphic principles. In several of the automated techniques t o
be described, biozonations and correlations will be based on average end
points of many local ranges. Figure 2.2 illustrates the concept of a n
average interval zone. Highest occurrences for two taxa (A and B) were
determined in nine sections (1-9). In most (7 out of 9) sections, the taxon A
exits above B. In two sections (numbered 3 and 9 in Fig. 2.2), B exits above
A. A variety of methods can be used t o estimate the average exit of taxon
A which occurs above the average exit of taxon B. Together these average
end points define an average interval zone.
Average interval zones can be combined with one another in order to
construct regional biozonations. Suppose that the eight exits in the

average
interval
zone

Fig. 2.2 RASC zonations are based on average stratigraphic events. The average interval zone between
the exits of taxa A and B begins before the highest occurrence of B in section 3 and ends before the
highest occurrenceof A in section 2.
 23

0.0 1

1T;
1-2
2 -3

1 .o 3-4
4-5
5-6
6-7

7-8

0.8 0.4 0.0

Distance

Fig. 2.3 Construction of dendrograrn for scaled highest occurrences of eight taxa. Intervals between
successive (average) exits are plotted along the distance scale of the dendrogram. Events which are close
together along the distance scale on the left (such as exits 3 to 6) form clusters which can be shaded in the
dendrogram. Clusters separated by longer distances can be useful as (RASC) zones in a regional
biozonation. Because average exits are used, events belonging to the same cluster are characterized by
more frequent cross-overs of tie-lines between sections.

0.0 -

1.0 -
i 6-8
8-1 0
10-12
12-1
Q 1-7
c 7-1 4
8 2.0-
U
14-1 6
b- 16-3
3-1 1

3.0 . 11-5
5-1 3
13-9
13
9-1 5
9
4.0.
0.8 0.4 0.0
Distance

Fig. 2.4 Same as Fig. 2.3 using lowest and highest occurrences to construct the dendrogram

example of Figure 2.3 are averages. The seven intervals between them
were plotted along the distance scale to the right and a dendrogram was
obtained by constructing perpendicular lines moving downward from the
points that represent the average interval zones. Each perpendicular line
24

ends when it meets the co-ordinate of an average interval zone. The

resulting dendrogram shows clusters for average exits that are close
together along the original distance scale. These clusters can be useful for
biostratigraphic correlation. An example of this technique using lowest
occurrences in addition t o highest occurrences is shown in Figure 2.4.
Zonations emphasize the temporal and spatial restriction of
morphologically distinct fossil taxa, arranged in zones. Good zonations
have zonal units with well-defined upper and lower limits, are easily
recognizable in many sections, correlate well and have been compared to
other regional or extra-regional zonations.
Correlation is one of the most widespread, abstract undertakings of
the mind and refers to causal linkage of present or past processes and
events. Such events can be inorganic, organic or abstract. Geological
correlation generally expresses the hypothesis that a mutual relation
exists between stratigraphic units. In a more narrow sense it means that
samples (or imaginary samples) from two separate rock sections occupy
the same level in the known sequence of stratigraphic events. Without
correlation, successions of strata or events in time derived in a specific
area would not contribute to our understanding of earth history elsewhere
(McLaren, 1978).
Suppose that the stratigraphic distribution of hundreds of taxa has
been sampled in dozens of wells or outcrop sections. Following a detailed
analysis, a range chart is proposed that synthesizes the information on all
ranges to arrive at total (maximum) ranges for each taxa. The range chart
is segmented, using co-existences of taxa and discrete taxon events, in
order to establish time-successive intervals. Each interval is called a zone.
When only last occurrences of fossils are known, such a chart portrays a
succession of events or partial ranges.
The critical and least understood step in the practice of correlation is
to actually tie the zones (back) to the individual sections. This may be a
difficult undertaking when the individual stratigraphic record shows
frequent inconsistencies due to sampling problems, reworking, unfilled
ranges because of facies changes, and other factors.
Ideally, the individual fossil record as observed in each rock section
should be compared to a regional standard prior to actual correlation.
Insight should be gained in the likelihood that observed events occur
where the standard (zonation) suggests that they should be found. In
 25

practice, the paleontologist will make a judgement on the outliers, or

events to be rejected or moved up or down in a section. Next, the
paleontologist will in each rock section define the successive zones in such
a manner that a minimum number of (key) taxa for each of the zones fall
outside the suggested zonal limits. Mismatch of the zones and the
individual record is explained as noise or strictly local correlation
character of the zones. Obviously, this is ideal terrain for a quantitative
approach where more than one solution can be proposed depending on
thresholds selected and where error bars may show uncertainty of
correlation and zonal limits.
Partially under the influence of a paleomagnetic reversal scale, which
promises virtually isochronous correlations for horizons in which a
paleomagnetic event has been unambiguously determined, efforts have
been made to establish detailed sequences of evolutionary fossil data. This
effort has been particularly successful in the siliceous and calcareous
marine plankton record of the last 150m.y., as preserved in Deep Sea
Drilling Program sites. In theory this allows for more or less reliable point
correlation in time, but in practice, independent corroboration using the
correlation of as many types of events as possible remains desirable.
In this vein, it is important t o establish the separation by necessity of
the reference framework of fossil taxa and rocks from abstract geological
time. Biostratigraphy, the global or regional record of paleontological
events or zones and their limits, used to correlate rock sequences, is the
common link between lithostratigraphy and chronostratigraphy.
Commonly it is assumed that correlation lines correspond to time lines,
but this remains a hypothesis (Drooger, 1974). To equate biostratigraphy
with chronostratigraphy and a priori substitute biozone for chronozone is
misleading. Although biostratigraphically perfect correlation can be
strongly diachronous, it may nevertheless be of value in sedimentary
basin analysis. The assumption of contemporaneity has to be verified
through other means, particularly by comparison t o correlations using a
particular zone elsewhere and through superposition of multiple
correlative units.
Chronostratigraphy, which has led t o the development of the
commonly used scale of geological stages, is essentially relative. As a
measure of relative age in geological history, reference is made t o the
standard chronostratigraphic scheme made up of successive stages like
Cenomanian, Turonian, Coniacian in the Cretaceous system. The stage
26
unit is a well-delimited body of rocks of a n assigned and historically
agreed upon relative age, younger than typical rocks of the next older
stage, and older than typical rocks of the next younger stage.
The accurate portrayal of geological history demands that relative
and subjective scales be modified into a numerical, linear scale. The
conversion of a relative to a so-called absolute scale, measured in units of
linear time like one million years is embodied in geochronology.
Numerous well-identified stratigraphic samples with accurate radiometric
age determinations are needed to calibrate the bio-magnetostratigraphic
scales in linear time.

2.3 Quantitative versus qualitative stratigraphy

In stratigraphy, there has been a considerable amount of discussion
regarding whether or not a probabilistic approach should be used. Harper
(1981) has stressed the need for a quantitative and statistical approach for
inferring succession of fossils in time. He has argued that most, if not all,
stratigraphic paleontologists make subjective assessments of t h e
probabilities of competing hypotheses regarding the ranges of taxa in
time. According to Harper (1981, p. 445), these assessments can and
should be backed up by quantitative methods and statistical tests. Others
(e.g. Jeletzky, 1965) have pointed out that quantitative methods either
explicitly or implicitly bring in new assumptions which could be too
restrictive. The greatest drawback of some types of quantitative methods
is that unequal things may be treated equally. Jeletzky (1985, p. 138)
based zonal schemes on index fossils replacing or completely ignoring a
great many other, facies-bound or long-ranging fossils often comprising
the bulk of the faunas concerned. A naive statistical approach based on
counts of all fossils would have led to inferior results.
It seems obvious that statistical methods are most useful in subfields
of paleontology which are rich in sampling points and taxa, especially if
use is made of standardized sampling methods and if valid conclusions
should be drawn by the elimination of “noise” for decision-making (e.g.
from micropaleontological information in oil exploration). The following
quotations from Schindewolf (1950, p. 79-80) as translated by Jeletzky
(1965, p. 139) for relation between quantitative “faunal” and qualitative
 27

“species zone” methods remain valid to-day as a summary for the relation
between quantitative and qualitative methods:
“It would seem to me that there is no need to make a choice here, that is, the two methods are not
usually exclusive but complementary. It is indeed not at all possible to draw a sharp boundary between
them. In order to achieve a greater precision in chronology, we use sometimes (in the case of species
zones), second or third series of species in addition to our principal evolutionary series of species. We
compare, furthermore, the time ranges of individual species with one another and so succeed in
recognition of a number of subzones. In such instances, one already considers a certain percentage of the
total fauna. This naturally constitutes a transition to the faunal method. In practice, the latter method
also does not ever utilize the sum total of forms available but only a selection therefrom. The long-
ranging, chronologically useless representatives of a fauna, which usually form its percentage wise
predominant element, are in this case quietly denied any consideration.”

“A community of organisms is a complex thing, the components of which are characterized by very
different behavior. Some of the individual forms (taxa) are extremely dependent on facies. They only
bloom under quite definite, narrowly limited conditions of life. If these conditions are altered, they
become extinct locally in some instances. In other instances, they emigrate and reappear sometimes, at
least in the instances of long-ranging species in considerably younger horizons, the conditions of
deposition of which have satisfied their specific bionomic requirements. Other organisms are less facies-
dependent. However, their sensitivety varies so that the individual forms concerned (taxa), in turn,
behave very differently whenever the conditions of life undergo changes. The changes of facies are
therefore apt to result in faunal discordances and strong variations in the composition of the faunas
concerned.”

Amongst quantitative stratigraphers, there has been discussion

about whether one should adopt a probabilistic or a non-probabilistic
(axiomatic, wholly deductive, or deterministic) approach. Harper (1981,
p. 442)has argued that a non-probabilistic approach may lead to relative
age hypotheses which should not be proposed because they are neither
falsifiable nor verifiable. As a starting point for discussion, Harper made
the following three assumptions:
1. The principle of superposition applies at any given sample site.
Owing to facies changes, the principle is best restricted, where possible, to
individual sites where superpositional order can actually be seen in
outcrop, or where it is obvious as in a borehole in a structurally simple
area.
2. The range of a taxon a t any given sample site has not been
extended upward by reworking (Jones, 1958;Wilson, 1964)or downward
by stratigraphic leaks (Jones, 1958; Foster, 1966). (In exploration
28

micropaleontology, one also has to avoid downward extension due to cave-

ins in wells.)

3. If two taxa occur together in a given narrow sample horizon (bed),

then their temporal ranges overlap i n geological time (Edwards, 1978,
p. 248).
Harper (1981, p. 443) remarked t h a t assumptions 1 and 2 a r e
essential to a non-probabilistic approach. Assumption 3 is expendable if
co-occurrences by themselves are not used to infer overlap. According to
Harper, there are 13 basic relative age hypotheses for any pair of taxa A
and B (Fig. 2.5). Hypotheses numbered 10A-B and 11A-B which assess
that the two taxa are sequential in time, may be falsified but not verified
using the three assumptions (1-3). Hypotheses 1-9taken individually can
neither be verified nor falsified. No single one of them can be verified
since any conceivable available data will be consistent with the other
eight. Harper (1981) concluded that a non-probabilistic approach of this
type is not fruitful. On the other hand, a probabilistic approach working

8P 1 OA I:, 1 1 5
It
Fig. 2.5 Possible relative age hypotheses for two taxa A and B according to Harper (1981). Vertical line
segments with arrows indicate ranges of taxa in time. Two hypotheses (10 and 11) are further divided on
the basis of presence or absence of a time gap between ranges of the two taxa.
 29

with preferred sequences rather than all individual sequences allows

significance tests that are based on a comparison between “sample” means
and hypothetical “population” means.

Fossils, taxa and events

From the previous discussions it is clear that in biostratigraphy
relatively little use is made of possible variables such as frequency of
individual fossils belonging t o a specific taxon; e.g. measured per sample
or per unit area of outcrop. To a large extent, the various types of
biostratigraphic zones are defined on presences and absences of taxa
rather than abundance data. The paleontologist looking for fossils in the
field commonly attempts to recognize as many different taxa as possible.
The ranges of these taxa are of special interest. The paleontologist usually
tries t o find the stratigraphically lowest as well as the highest occurrence
of each taxon within a section (local range) or region. In general, it is more
efficient t o recognize among the hundreds or thousands of fossils the
presence of one or more fossils belonging to a specific taxon, rather than to
attempt to classify and count all individual fossils. It will be discussed in
Chapter 3 that microfossil abundance data can be useful for correlation in
biostratigraphy. However, very large samples and much effort may be
required to obtain fossil abundance data which are relatively precise. It is
more effective t o establish the presence or absence of a taxon, because, in
general, more information is provided by presence-absence data of many
taxa than by precise abundance data for relatively few taxa.
Nevertheless, the presence of a taxon in a bed is determined by its
abundance in this bed. This abundance reflects the chances that the taxon
occurred at a given place, became fossilized, was found and correctly
identified, which in themselves reflect hit-or-miss processes. It will be
seen that when quantitative correlation of the presence-absence data for
taxa in different stratigraphic sections is attempted, this effort is
commonly hampered by existence of numerous inconsistencies which must
be resolved before meaningful correlation is possible.
The quantitative analysis of abundance data can be useful in specific
subfields of paleontology such as palynology. For example, Christopher
(1978) successfully performed pairwise comparison of time series for
30
quantitative palynologic correlation of Upper Cretaceous sections from the
Atlantic coastal plain.

2.4 Local versus regional ranges of taxa

Each fossil taxon has a lowest and a highest occurrence in the local
range for a continuous outcrop section or a single well, as well as in the
regional composite range for a number of stratigraphic sections. A
regionally-based range chart is more useful for stratigraphic correlation
than the local ranges showing superpositional relations that often are
mutually inconsistent. The positions of highest occurrences for a regional
range chart commonly are underestimated, and those of lowest
occurrences overestimated when distances t o observed ends are measured
from the base of each stratigraphic section upward and averaged between
sections. This problem will be discussed at length in the next section.
Suppose, however, that this type of bias can be neglected and that it has
been possible to measure the local ranges for a number of taxa in a number
of sections. Then combining sections with one another t o construct a
single range chart may give misleading results for a number of other
reasons. The problem was illustrated by Davaud (1982) as follows.
Figure 2.6 is a theoretical example showing distribution in space and
time of 7 different taxa and their true chronological succession. Obviously,
the local ranges in the four sections A-D differ from the true regional
succession of the biological events. Differential preservation of the taxa
during fossilization may create further differences between local and
regional ranges. So do the processes of sedimentation, compaction, and
other processes. Figure 2.7 illustrates possible influence of differential
sedimentation on the ranges for a single species. Disregarding other
factors, a combination of the living range factor (Fig. 2.6) and the
differential sedimentation factor (Fig. 2.7) resulted in the sedimentary
record of Figure 2.8. Obviously, the local ranges of Figure 2.8 do not
provide good estimates of the local ranges in Figure 2.6. Neither can a
composite range chart based on Figure 2.8 provide an approximation to the
chronological succession of “biological” events in Figure 2.6.
Fortunately, it generally is possible in practice to design experiments
in order t o check whether or not the factors illustrated in Figures 2.6 to 2.8
have significant effects. For example, differences in living range can be
evaluated by performing separate data analyses on subsets of a regional
 31

Fig. 2.6 Theoretical example of Davaud (1982)showing distribution in space and time of seven different
taxa with true chronological succession.

database (cf. Section 4.7). These subsets which correspond t o geographical

subregions would yield different results if there were large shifts in the
living ranges of the taxa. It also may be possible t o evaluate this factor by
means of multivariate analysis using the geographical locations of the
stratigraphic sections as variables (cf. Section 2.4). The influence of
differences in rates of sedimentation between stratigraphic sections can be
evaluated if sufficient information is available t o establish the sediment
accumulation histories for individual sections using the numerical
geological time scale (see Chapter 9).

2.5 Estimation of the highest and lowest occurrences of taxa

Figure 2.9 illustrates the relationship between fossil finds, ends of
observed local range and “true” ends of the local range of a taxon. In
recent years, several methods have been developed for estimating the
“true” highest and lowest occurrences of a taxon (Jasko, 1984; Springer
and Lilje, 1988;Strauss and Sadler, 1989). This type of estimation is only
possible if simplifying assumptions are made, e.g. constant facies with
32

Space Space Space

la) (bl It1

Fig. 2.7 Diagrams to illustrate how biological events are recorded in sediments (after Davaud, 1982).
Diagram (a) shows time-space domain for a particular species. Population density is reflected by points
density. Diagram (b) illustrates that during same period of time and in same geographic area, the
sedimentation rate changed. When the sedimentation rate is applied to points of diagram (a) and
integrated over time, the points are moved to new positions in the sedimentary record as shown in
diagram (c). If the probability of detection is proportional to density of points in the sedimentary record,
the end point of the chronological range of a species could be underestimated, especially if sedimentation
rate was high at time of biological disappearance of the species.

(A) (D)
-
5
21 4
I
?
I
I 1

I 1 I
I1
3
I

T I
d I5
?I
Fig, 2.8 Sedimentary record of biological events in four stratigraphic sections corresponding to the
theoretical example of Fig. 2.6. Distortion due to differential role of sedimentation was similar to the one
shown in Fig. 2.7 (b).

constant average rate of sedimentation. Figure 2.10 (from Strauss and

Sadler, 1989) shows local ammonite ranges in late Cretaceous strata of
Seymour Island, Antarctic Peninsula. The observed local ranges and finds
are from Macellari (1986). The highest occurrences were obtained by
 33

c
”true“
range ;:li
f -e,-
I

base
observed
range

Fig. 2.9 Relationship between observed range extending from time t l to t ~and , “true” range extending
from time 81 to 82. Strauss and Sadler (1989) assumed that the probability of finding a fossil is constant
across its true range. If a species was less abundant at its time of appearance or disappearance, a s
illustrated by the density curve in the diagram, it becomes more difficult to estimate the true range even
if facies and sedimentation remained constant.

Strauss and Sadler as unbiased point estimators and their upper range
extension to 95 percent confidence interval. These authors used the
Dirichlet distribution which results from a Poisson process for uniform
sedimentation. It was assumed that each fossil existed for an unknown
period of time. The chances of finding it remained equal during this
period. The density curve for highest finds has a tail that extends in the
stratigraphically downward direction under these conditions.

Jasko (1984)used a different model to estimate precision of the

observed lowest occurrence of a taxon. He assumed that initially the
population of a taxon increases its size exponentially as established e.g. for
bacterial colonies in the laboratory. The average number of specimens per
unit volume would follow a Poisson distribution. The combination of these
two distributions leads t o a new (compound Poisson) frequency
distribution permitting estimation of the average range ( r ) and its
standard deviation ( d ) for a given number of specimens (see Table 2.1). In
practice, it may be possible t o determine the local range from the
observations (see Table 2.2) and to set it equal t o the average range. The
corresponding standard deviation then expresses the uncertainty in the
position of the lowest occurrence. In the example of Table 2.2, the
compound Poisson distribution provides a good fit from 2700 f t downward.
34

Fig. 2.10 Ammonite ranges in late Cretaceous strata of Seymour Island, Antarctic Peninsula. Observed
local ranges (heavy vertical lines) and actual finds (solid circles) after Macellari (1986, Fig. 5).
Extrapolated end-points of ranges according to Strauss and Sadler (1989, Fig. 1). Light vertical lines
represent upper range extensions to unbiased point estimators. Dashed vertical lines a r e upper range
extensions to 95 percent confidence intervals. Numbers assigned to taxa a r e a s follows: 0 =
Diplomoceras lambi; 1 = Maorites seymourianus; 2 = Kitchinites darwini; 3 = Grossouurites gemmatus;
4 = Maorites weddelliensis; 5 = M. densicostatus morphotype-alpha; 6 = Kitchinites laurae; 7 =
Anagaudryceras seymouriense; 8 = Maorites densicostatus morphotype-gamma; 9 = Pachydiscus
riccardi; 10 = Maorites densicostatus morphotype-beta; 1 I = Pseudophyllites loryi; 12 = Pachydiscus
ultimus.

This is indicated by t h e close correspondence between observed

frequencies and expected frequencies based on the statistical model. In
total, 25 microfossil forms were observed for the bottom 3 classes in Table
2.2. The ratio of standard deviation to range is 0.348 if n=25. Because the
lowest occurrence was observed in a sample a t 3446 ft., the local range is
3446-2700 = 746 ft. The standard deviation for the lowest occurrence is
estimated to be 0.348 X 746 = 260 ft. If the position of the lowest occurrence
would be normally distributed (i.e. satisfying the Gaussian curve model),
there would be a 95% probability that the true lowest occurrence is below
+
3446 1.645 X 260 = 3874 ft.
 35

TABLE 2.1

Averages ( r ) ,standard deviation (d)and their ratio ( V = d / r ) as functions of sample size ( n ) as obtained
by means of computer simulation experiments (after Jasko, 1984).

n r d V n r d V
I oon 985 16 3 Ill 1259 405
2 864 1093 1265 17 3203 1259 393
3 I355 1 I28 832 I8 3231 1247 386
4 1663 I I63 699 19 3285 1263 385
5 I910 I I91 623 20 3323 1273 383
6 2 112 I188 562 21 3370 I267 376
7 2263 I199 530 22 3432 I288 375
8 2412 I209 501 23 3514 1270 361
9 2541 I206 475 24 3534 I277 361
10 2638 I227 465 25 3586 I249 348
II 2737 I247 456 26 3 563 I276 358
12 2817 I237 439 27 3648 I287 353
13 2893 1250 432 28 3692 I272 345
14 2971 I 250 421 29 3698 I 269 345
15 3 052 I 254 411 30 3777 I 292 342

Possible models for the shape of the frequency distribution for

positions of highest and lowest occurrences will be discussed in the next
section. It is noted here that Strauss and Sadler's model for highest
occurrences implies t h a t t h i s distribution is not symmetrical.
Theoretically, in their model, the last find has a distribution with a longer
tail in the stratigraphically downward direction. Instead of this, the
distribution of Strauss and Sadler's estimated end of the range has a long
narrow tail that extends upwards, especially for fossils with relative few
finds such as Maorites weddelliensis (4) and Pseudophyllites loryi (11) in
Fig. 2.10. Jasko's model for lowest occurrences (Table 2.2) implies an
asymmetrical frequency distribution with its long narrow tail extending
downward. The estimated lowest occurrence is skewed in the same
direction. Thus the 95% confidence limit of 3874 ft for the lowest
occurrence estimated in the preceding paragraph is probably incorrect
because it was based on the symmetric Gaussian distribution model. If
Jasko's model is correct, the 95% confidence limit has a depth value
greater than 3874 ft.
A third model for sampling bias resulting in artificial range
truncation was developed by Signor and Lipps (1982). These authors deal
with the phenomenon that taxa begin to disappear from the fossil record
before mass extinctions actually take place. Figure 2.11 illustrates this
idea. The line in Figure 2.11A represents a n abrupt change in the
diversity of various taxa coinciding with mass extinction (e.g. a t the
36

TABLE 2.2

Jasko's (1984) example of frequency ( = number of specimens) of a microfossil species in a borehole

section. Lowest occurrence in sample a t 3446 ft.

Depth interval in ft Actual frequency Expected frequency

2100 - 2400 41 40.1
2400 - 2700 26 23.6
2700 - 3000 11 13.9
3000 - 3300 9 8.2
3300 - 3600 5 4.8

A B C

time time time

Fig. 2.11 Model of Signor and Lipps (1982) for alteration of diversity patterns by artificial range
truncation. In Fig. 2.11A, diversity is suddenly reduced by a catastrophic extinction event. Imposing the
artificial range truncation model illustrated in Fig. 2.118 on the pattern of Fig. 2.11A produces the
apparent gradual decline in diversity of Fig. 2.11C.

Cretaceous-Tertiary boundary). Figure 2.1 1B plots a n arbitrary

probability curve giving the probabilities of different degrees of range
truncation. This produces the apparent diversity curve shown in Figure
2.11C. Note that the slope of the hypothetical curve in Figure 2.6B
continues to increase until the time of the mass extinction. Different
sedimentary sections would be characterized by different curves. For
example, if the curve of Figure 2.11B is representative for nearshore
marine and terrestrial sections, the deep sea plankton record would have a
curve whose slope increases less initially and becomes steeper near the
time of the mass extinction (Signor and Lipps, 1982, p. 294). Thus the
apparent diversity curve for oceanic microplankton is closer to actual
 37

diversity than e.g. the curve for dinosaurs below the Cretaceous-Tertiary
boundary (cf. Russell, 1975,1977; Van Valen and Sloan, 1977).

2.6 The frequency distributions of highest and lowest occurrences

of t a x a
Figure 2.12 shows a hypothetical relationship between relative
abundance, observed highest occurrence and relative time for two taxa.
Agterberg and Nel (1982b) introduced this example t o illustrate that the
abundance of a taxon may have changed through time. The range of the
frequency curve of its observed highest occurrence is narrower than the
range of the abundance curve although these two curves end at the same
value along the time axis. Especially if a systematic sampling procedure is
carried out such as obtaining cuttings at a regular interval (e.g. 30 ft or 10
m) along a well in exploratory drilling, the highest occurrences of two taxa
with overlapping frequency curves may be observed to be coeval. The fact
that two taxa have observed highest occurrences in the same sample does
not necessarily mean that they disappeared at the same time. Rare taxa
such as taxon B in Figure 2.12 are likely to have wider ranges for their
highest occurrences.

OBSERVED
HIGHEST
/ OCCURRENCE

R E L A T I V E T I M E SCALE

Fig. 2.12 Schematic diagram representing frequency distributions for relative abundance (broken lines)
and location of observed highest occurrence (solid lines) for two taxa. Vertical line illustrates that
observed highest occurrences of two taxa can be coeval even when the frequency distributions of these
two taxa are different.
38

z z
0
0
F
+
V
3 z
I-
X
W

I
I
I
REWORKING
OOWNHOLE I M ISIDENTIFICATION
; REWORKING

TIME OR ROCK THICKNESS

(01

I I
I I
I
a
l
I
I
I
I
CONTAMINATION,’ MISIDENTIFICATION

TIME OR ROCK THICKNESS

(b)

Fig. 2.13 Edwards’ (1982a) model to display probability of observing lowest - or highest-occurrence event
relative to “true” time of evolution or extinction in outcrop or core material for (a) first occurrence event;
and (b) last occurrence event. According to Edwards (1982), details for curves will vary for every
individual taxon, and gross shapes of curves will vary with kind of organism (e.g. rapidity of dispersal,
facies control) and nature of sample material (core, outcrop, cuttings).

Figure 2.12 shows symmetrical, “normal” curves for the observed

highest occurrences. It can be assumed that, in reality, these curves are
not symmetric but skewed. Figure 2.13 (from Edwards, 1982a) is a n
attempt a t displaying asymmetric curves for lowest a n d highest
occurrences along with the main factors controlling the shapes. It is noted
however, that Edwards’ assumption on the nature of the skewness differs
from t h a t implied by Jasko’s model, in which the tail of observed lowest
occurrences extends i n the stratigraphically downward direction ( I n
Edwards’ model it extends upward). In the model of Strauss and Sadler,
the tail for highest occurrences points downward which is i n agreement
 39

with Edwards’ assumption. Likewise, the model of Signor and Lipps (Fig.
2.11B) is i n agreement with t h a t of Edwards because the slope of their
curve continues to increase in the stratigraphically upward direction.

Figure 2.14 from Baumgartner (1986) also supports the model of

Edwards (Fig. 2.13). I t is illustrated in this diagram why a composite
range based on many sections generally is relatively short ( = iAB)when i t
is based on mean positions of the frequency distributions for highest and
lowest occurrences. In the Unitary Associations method, stratigraphic
correlation is based on the three zones i n the column on the right of Figure
2.14. The range of taxon A extends higher than the interval eAand t h a t of
Taxon B occurs below eB. The latter two intervals are based on the
symmetrical Gaussian curves. A curve of this type has the property that
68 percent of the observations deviate less than one standard deviation
from its mean. If eA and eB would be extended to points located two
standard deviations from their mean, t h e probabilistic range c h a r t
becomes approximately equal to the zonation resulting from the Unitary
Associations method. These wider probabilistic ranges would contain
approximately 95 percent of the observations.

Arrorlatlonr

bases
species E

A B

tops
species A

A A C D

Fig. 2.14 Baumgartner’s (1986) model for frequency curves of last appearance of species A and first
appearance of species B. The two species are actually co-occurring in section 7. The asymmetrical
smoothed curves in Fig. 2.14C a r e based on the bar-graphs representing the observed frequencies of Fig.
2.14B. In a probabilistic model, it could be assumed that these curves are symmetrical (broken lines)
extending upward and downward from the mean positions. If the means a r e used for constructing a
range, the result is ~ A B . A symmetrical Gaussian curve has the property that 68 percent of the area
undder the curve is contained between its inflection points located a t the mean plus or minus one
standard deviation. These intervals a r e shown as eA and eg. The Unitary Associations method would
result in the overlapping ranges for species A and B shown in Fig. 2.14D.The latter result would also be
obtained by using the Gaussian curves and assuming that and eg would extend two instead of one
standard deviations on either side of the mean.
40

Edwards (198213) has pointed out that if both highest and lowest
occurrences of taxa are used, there is a possibility that in some methods of
ranking, the highest occurrence of a taxon would end up below its lowest
occurrence. Possible and impossible arrangements for the events resulting
from 2 taxa are shown in Figure 2.15. Note t h a t all impossible
arrangements have in common that either A (lowest occurrence of first
species) occurs above B (highest occurrence of first species) or that C occurs
below D for the second species. If in a statistical method all events were t o
be treated independently, the final ranking might contain impossible
arrangements. A problem of this type can be avoided, e.g. by recognizing
during the coding of the stratigraphic events or within the computer
program for statistical analysis, that the lowest occurrence is below the
highest occurrence for each taxon in theory and practice.

D c l B A 1
C I D D

:I : IT
IVPOSSIBLE
A B T
IVPOSSIBLE

: 11
A
A
D
1

B r
IMPOSSIBLE
B T
IMPOSSIBLE
' I 1 :TI
C
IMPOSSIBLE

C " B 1 tLl
A I T 1,
IMPOSSIBLE IVPOSSIBLE
TT
IWOSSIBLE

D
:
C 11
c
::I
"
F
A 11 A
C
B
'I
D T
IVPOSSIBLE
b T
IFIPOSSIBLE
D TT
IVPOSSIBLE

A I' : C
A
B11
"
C
B T
IMPOSSIBLE
D
B T T D T : TI
IMPOSSIBLE IMPOSSIBLE IVPOSSIBLE

A A
B
I" B
A '
!il
D T

IMPOSSIBLE
D TT
IVPOSSIBLE IMPOSSIBLE

Fig. 2.15 The 24 arrangements of 4 events, where A and B are first and last occurrences of one species,
and events C and D are first and last occurrences of a second species. Only 6 of these arrangements are
possible (from Edwards, 198213). Quantitative stratigraphers should always look for impossible
arrangements in computer output and modify their algorithm if required.
 41
Several possible frequency distribution models for highest and lowest
occurrences are shown in Figures 2.16 and 2.17. The spike (A) represents
abrupt disappearance of a taxon in Figure 2.16 and its immediate
widespread appearance in Figure 2.17. Because the spike is symmetrical,
the frequency curve also must be symmetrical when it is narrow (possibly
B in Figs. 2.16 and 2.17). Wider frequency curves have different values for
their mode (l),median (2) and mean (3), respectively. Curves for which
the order of the mode, median and mean is 123 are positively skew in the
direction of time. Those with order 321 are negatively skew.
Symmetrical curves have coinciding mode, median and mode. As
shown in the captions of Figures 2.16 and 2.17, all models discussed so far
correspond t o one of the 12 possibilities. It can be assumed that, with the
possible exceptions of A and C in Figures 2.16 and 2.17, all these frequency
curves exist in the fossil record. In practice, it is almost always impossible
t o precisely measure the shapes of the frequency distributions of the
highest and lowest occurrences of a taxon because one would need large
numbers of sections that are calibrated precisely according to time-lines.

Fig. 2.16 Six possible shapes for the frequency distribution of the observed last occurrence of a taxon. the
top (t) is the truly last occurrence. The numbers 1, 2 and 3 represent mode, median and mean,
respectively. These three statistics coincide for a symmetrical curve. Most paleontologists assume that
Fig. 2.16D is the most widespread shape. Arrow points in direction of time.
42

A C

E F
123

Fig. 2.17 Six possible shapes for the frequency distribution of the observed first occurrence of a taxon.
The base (b) is the truly first occurrence. The numbers 1, 2 and 3 represent mode, median and mean,
respectively. Opinions are divided as to which shape (Dor F) is most widespread.

The subject of shapes of frequency distributions of highest and lowest

occurrences largely remains in the realm of speculation, as is indicated by
the fact that no concensus has been reached in literature. It seems that, in
the absence of outliers due to reworking and other disturbing factors, the
majority of paleontologists assume the shape of Figure 2.16D for the
frequency distribution of the tops and that of Figure 2.17F for the bases.
Both distributions have their longest tail in the stratigraphically
downward direction. Figure 2.17F as the preferred model for first
appearance data is contrary t o the models of most quantitative
stratigraphers (see before). However, as pointed out by Shaw (1964, p. 94),
many paleontologists assume that there is a period (Shaw’s “hemera”) in
the history of any species before it reaches its acme (Shaw’s “epibole”) in
terms of numbers of individuals. Such a model is most likely to result in
the shape of Figure 2.17F. Later in this book (see Chapter 91, a method
will be discussed for actually measuring the skewness of the frequency
distributions of bases and tops. However, the number of applications of
this method remains t o o small t o decide which models are most
widespread.
 43

lhl

Fig. 2.18 Examples of the effect of averaging illustrate the central limit theorem of mathematical
statistics. No matter what shape the frequency distribution of the original observations (a), taking the
average of two (b), four (c) or 25 (d) observations not only decreases the variance but brings the curve
closer to the normal (or Gaussian) limit (after Lapin, 1982; and Davis, 1986).

In the RASC method of ranking and scaling, the initial objective is t o

estimate the mean value (3 in Figs. 2.16 and 2.17) of the highest and
lowest occurrences as precisely as possible. Biozonations as well as
stratigraphic correlations are based on these mean values. The advantage
of this procedure is that the mean can be precisely estimated regardless of
the shapes of the frequency distributions of the events. This relative
independence of shape is due to the central limit theorem of mathematical
statistics (see Fig. 2.18) which states that addition or averaging of n
independent random variables gives new random variables that become
normally distributed when n increases. In the scaling part of RASC,
distances between successive mean event locations are estimated by
averaging many indirect distance estimates. Each of the latter estimates
is a value originating from a frequency distribution that itself is a n
average of the frequency distributions for three separate stratigraphic
events. Although the shapes of the original distributions may not be
normal, the resulting frequency distributions based on sets of three events
44

L i XL FT Vl iFi I( T flF

Fig. 2.19 Frequency histograms for finding a taxon within its range before and after mixing (from
Edwards, 1982b).See text for further explanation.
liltl

are probably approximately normal. Further averaging of many indirect

estimates yields mean event locations along the RASC scale that can be
very precise.
Ranges based on mean positions are shorter than ranges resulting
from attempts to estimate the locations of the true tops and bases ( t and b )
in Figures 2.16 and 2.17. Such maximal ranges attempt to represent the
periods of time that taxa existed in a region. Estimation of the true end
points is more difficult than estimating the mean event locations for
several reasons: (1) statistically, the largest or smallest value in a sample
of n values drawn from a population has a standard deviation which is
greater than that of the mean of all values; and (2) the influence of
“outside” values not belonging to the statistical population on the average
range is much smaller than their influence on the maximal range. This is
because maximal ranges would be based on values due to outside factors
such as misidentification, contamination, downhole caving or reworking
(cf. Fig. 2.13) unless these factors can be identified with certainty so that
all outside values can be eliminated.
 45

It is possible that the shape of the frequency distribution is changed

because of one or more outside factors. Berger and Heath (1968) proposed
a model for postdepositional mixing which was used by Edwards (1982) in
computer simulation experiments. Figures 2.19 shows results for two
initial distributions (A) and (B) after variable amounts of mixing (to
degrees 1,2 and 3). Degree 1 (LIM = 4) mixing led t o a downward shift of
the modes as shown in the resulting frequency curves (C)and (D). The
effect of increased mixing t o degrees 2 (LIM = 2) and 3 (LIM = 1)is shown
in (E) and (F) for the second initial distribution only. Edwards (1982b)
used the formula P = Po exp (-LIM) of Berger and Heath (1968) where Po
and P represent the probability of finding the taxon within its range before
and after mixing, respectively; L is the sample interval, and M is the
thickness of the zone of mixing. The tail on the right (in direction of time)
is increasing in length and the end product after mixing becomes nearly
symmetrical in Figure 2.19F.
This Page Intentionally Left Blank
 47

CHAPTER 3
APPLICATIONS OF MATHEMATICAL STATISTICS AND
COMPUTER SCIENCE TO ZONATION,
CORRELATION AND AGE INTERPOLATION

3.1 Introduction
This chapter contains background information f o r various
applications of mathematical statistics and computer science. It can be
skipped by readers who are not primarily interested in mathematically-
based theory. Concepts and methods t o be discussed include:
(1) probabilities, Bernoulli trials and the binomial model; (2) graph theory;
(3) multivariate analysis; (4) method of maximum likelihood; and
( 5 ) smoothing splines. Most of these techniques are illustrated by means of
geological examples of interest in paleontology and stratigraphy although
the emphasis in this chapter is on mathematical background. Not all
mathematical discussions are contained in this chapter. Other techniques
will be introduced in separate sections within later chapters as needed.
Modern mathematics and the theory of probability and statistics are
formally based on set theory. There have been several interesting
attempts t o formulate conventional stratigraphy in strict logico-
mathematical terms (Dienes, 1974; 1982; Dienes and Mann, 1977;
Carimati et al., 1982). The language of set theory, although a necessity in
pure mathematics, is not of immediate practical usefulness in stratigraphy
which has a well-developed language of its own. Although superpositional
relations between stratigraphic events can be precisely formulated in
terms of sets, the nomenclature of set theory is unpalatable t o most
stratigraphers as pointed out by Tipper (1989, p. 480).
The mathematical techniques introduced in this chapter are required
for statistical applications and for use in computer-based graphs and
graphics. Although these techniques are widely applied in other fields of
science, and may be elementary to those trained in mathematical
statistics, they have been used hardly at all in stratigraphy. The purpose
of this chapter is not only to review statistical methods that have been
48

applied in stratigraphy, but also t o show t h a t other methods (e.g.

maximum likelihood method) can be used to refine existing methodologies.

3.2 Binomial test for randomness

The binomial test for randomness will be briefly discussed (cf. Hay,
1972; Southam et al., 1975; Blank and Ellis, 1982). If the sequence of a
pair of biostratigraphic events is random, the probability of one event
preceding the other is p = 1/2. Each observed superpositional relation is
thought to be the outcome of a Bernoulli trial. Suppose that two events (A
and B) both occur in N sections. Then the probability that A occurs above
B k times satisfies

P ( k ) = NCk2 - N
(3.1)

with the binomial coefficient being

[
NCk = N! k ! ( N - k ) !
I -l (3.2)

For example, if N = 5, then P(O)= P(5)= 1/32; P(1)= P(4)= 5/32; and
P(2)= P(3)= 10/32. These probabilities add to one. It is also possible t o
write P(0 or 5) = 1/16, P(1 or 4) = 5/16 and P(2 or 3) = 10/16. In practice,
the observation that A occurs k times above B generally cannot be
distinguished from B occurring k times above A when the hypothesis
p = E W N ) = 112 is being tested. In this expression, E( ...I denotes
expected value. K denotes the binomial random variable with observed
frequencies k (=O, 1, 2, ..., N). The test hypothesis obviously cannot be
rejected if KIN becomes equal to 1/2, a situation which may be observed
when N is even. For k > N/2, the probability
N
Pc(k) = 2 1 NCk2-N (3.3)
r=k

may be computed where the subscript c denotes that this probability is

c u m u l a t i v e . For t h e p r e c e d i n g e x a m p l e , P c ( 5 ) = 1 / 1 6 ,
P,(4) = 1/16 + 5/16 = 6/16, and PJ3) = 6/16 + 10/16 = 1. This
probability was tabulated by Hay (1972, Table 1 on p. 264). Next a level of
 49

significance (e.g. a = 0.05) can be selected. Then the hypothesis p = 1/2

will be rejected only if P,(h) C a.
The binomial test is useful when only two events are being compared
t o each other. If many events are to be considered simultaneously while
most values of N are small, this approach is less useful. For example, in
Figure 4.2 of Chapter 4 (see later), event A occurs 4 times above event C .
According t o the binomial test PJ4) = 1/8 = 0.125 for N = 4. This
exceedsa = 0.05 and the hypothesis that events 1 and 10 are coeval
( p = 1/2) therefore may not be rejected. Strictly speaking, it would have t o
be accepted . On the other hand, event A is separated from event < by 4
intermediate levels with other events in 3 of the 4 sections considered.
This would suggests that event A probably occurs above event < .
A multivariate statistical approach would be needed to test whether
or not two events are coeval when observations on many other events also
are available. Later, an approach (scaling method) will be developed which
permits the use of significance tests in which all events can be considered
simultaneously.

3.3 Binomial distribution model for microfossil abundance data

This section deals with statistical analysis of microfossil abundance
data. The microfossil record of the Portugese Oxfordian black shales
(Stam, 1986; Agterberg et al., 1990) will be used for example. In this case
history study it will be investigated whether, and t o what extent,
foraminifera1 abundance data can be used for detailed biostratigraphic
correlation in two sections of the black shale in the Montejunto area of
central Portugal.
In general, most biostratigraphic correlation is based on biozonations
derived from range charts using highest and lowest occurrences of species.
For example, in exploratory drilling a sequence of samples along a well in
the stratigraphically downward direction is systematically checked for
first occurrences of new species. The probability of rejecting a species in a
single sample depends primarily on its abundance. As a measure, relative
abundance (to be written asp) of a species in a population of microfossils is
commonly used. Together with sample size ( N ) ,p specifies the probability
of the binomial distribution with general equation:
50

P ( K = k ) = P ( k ) = NCk p k ( l - p ) N - k ( k = O , 1, ...,N
(3.4)

which represents the probability that k microfossils of the taxon with

relative abundance p will be found in a sample of N microfossils. Note that
for p = 1-p= 0.5, this probability reduces t o the one used in the binomial
test for randomness (Eq. 3.1). If p is very small, the binomial probability
can be approximated by the probability of the Poisson distribution.

P ( k ) = e-’Ak/k! ( k = 0 , 1 , ...,N)
(3.5)

which is determined by a single parameter A. The Poisson distribution can

be derived from the binomial distribution by keeping X = N p constant and
letting N tend t o infinity while p tends to zero.

The expected (or mean) value for a binomial distribution is E(K)=N p

and for a Poisson distribution: E(K)=A. The variance 0 2 M ) of the
binomial distribution is N p ( 1-p) while the variance of the Poisson
distribution satisfies 0 2 ( K )= E(K)= A .

Figure 3.1 (after Dennison and Hay, 1967) shows probability of

failure t o detect a given species for different values of p as a function of
sample size ( = N ) . For example, in a sample of N = 2 0 0 microfossils, a
species with p = 1 percent has probability of about 15 percent of not being
detected. This implies that the chances that one or more individuals
belonging to the species will be found are good. Unless its relative
abundance is small, the first occurrence of a species in a sequence of
samples can be established relatively quickly and precisely.

It is noted that the two scales in Figure 3.1 are logarithmic and that
the lines are approximately straight unless p is relatively large. This is
because the equation for zero probability of the Poisson distribution, which
provides a good approximation when p is small, plots as a straight line on
logarithmic graph paper. If 10 is used as the base of the logarithms, the
equation of each line in Figure 3.1 is simply loglo N=loglo A - loglo p with
P = P ( K = 0) = exp (-A) as follows from Equation (3.5).

The binomial distribution model on which Figure 3.1 is based also can
be used to estimate confidence intervals for any specific proportion value
( p ) . Unfortunately, it turns out that large samples would be needed to
estimate, with precision, the relative abundances of many different
species. In general, proportions estimated from actual samples are
 51

Fig. 3.1 Size of random sample (n)needed to detect a species occurring with proportional abundance ( p )
in population with probability of failure to detect its presence fixed at P (after Dennison and Hay, 1967).

uncertain. Moreover, the use of the binomial distribution model is based

on the assumption that the underlying population is a homogeneous
random mixture. This condition may hold true only locally, at the precise
place where a sample was actually taken. The proportions of the species
may change parallel and, in general more rapidly, perpendicular t o
bedding. It is hard to establish such changes because of the uncertainty in
the estimated values.
For these reasons, it is hazardous to use measured proportion values
for biostratigraphic correlation although it will be shown in the following
case history study that some species (e.g. Epistomina mosquensis) can be
useful for this purpose. The precision of proportion values also has been
studied in detail by palynologists. Maher (1972) h a s published
52

nomograms for computing 0.95 confidence limits of pollen data. A related

topic is t o study the precision of microfossil concentration measurements
by employing samples spiked with marker grains (Maher, 1981; White,
1990).

Geological background

Both syn-rift fault tectonics and changes in eustatic sealevel

influenced Jurassic carbonate through clastics marine sedimentation in
the Montejunto Basin, Portugal (cf. Stam, 1986; Agterberg et al., 1990).

Tojeira 1 Tojelra 2

\-
25

23
22
Metres
20

18

16

14

12
-9
11 -
-7
10

8
6 -6
6A
-5
5 -

3A -3
6.2 -12.1

Sandstone Shale Limestone

GSC

Fig. 3.2 Left side: Tojeira 1 section with sample members 6.2-6.29 (after Stam, 1986); ammonite zones
(Planula and Platynota Zones) of Mouterde et al. (1973) also are shown. This section is immediately
overlain by the poorly exposed sandy Cabrito Formation. Right side: Tojeira 2 section with sample
numbers 12.1-12.11 and 11.1-11.23(after Stam, 1986).
 53

Bathonian through Callovian carbonate bank and shelf apparently

became emergent in latest Callovian time due to widespread uplift or
sealevel fall. Renewed transgression in Middle Oxfordian led t o
bituminous algal and micritic t o oolithic limestones of the Cabacos
Formation, changing upward into thick-bedded micritic brachiopod
biostromes of the Montejunto Formation. Rapid deepening in latest
Oxfordian t o early Kimmeridgian time, when conditions became more
humid, led to sedimentation of dark grey shales of the Tojeira Formation,
followed upward by massive terrigenous-clastic fill (Cabrito and Abadia
Formations).

In Oxfordian time (approximately 150 Ma ago), at the onset of the late

Jurassic, a transition from one sedimentary mega-sequence into another
one took place. For example, in the North Sea Basin, the Lusitanian Basin
and the southern margin of Tethys ocean, now occupying the belt between
the central Himalayans and Tibet, the Oxfordian saw the sudden onset of
black shale deposition lasting up t o 15 Ma or more. Climate must have
become more humid; the black shale facies was probably also related t o
regional basinal deepening, in the absence of major relief rejuvenation
that would induce terrigenous clastic supply. In places, the shales
constitute major hydrocarbon source rock.

Location of Tojeira sections; summary of Stam’s quantitative results

The Lusitanian Basin originated in the late Triassic - early Jurassic
as a result of movements along Hercynian basement faults including the
prominent Nazare strike slip fault. Several cross-sections i n t h e
Montejunto area were sampled by Stam (1986) for quantitative analysis of
Middle and Late Jurassic Foraminifera in Portugal and its implications
for the Grand Banks of Newfoundland. The so-called Tojeira 1 section
with sample numbers 6.2-6.29 (after Stam, 1986) is shown in Figure 3.2
(left side). It is continuously exposed and occurs about 2km southeast of
the Tojeira 2 section (Figure 3.2, right side) with Stam’s sample numbers
12.1-12.11 and 11.1-11.23. The Tojeira 2 section is not continuously
exposed; two missing parts are estimated to be equivalent to 35m and 50m
in the stratigraphic direction, respectively.

Tojeira shales contain a rich and diversified (over 45 taxa) planktonic

and benthonic foraminifera1 fauna, including Epistomina mosquensis,
E. uhligi, E . volgensis, Pseudolamarckina rjasanensis, Lenticulina
54
quenstedti, and Globuligerina oxfordiana. Stam determined from 21 t o 43
species per sample in Tojeira 1; between 301 and 916 benthos was counted
per sample; proportions were estimated f o r 14 species. The
plankton/benthos (P/B) ratio also was determined for each sample.
Correlation coefficients for relative abundance estimates of the benthonic
Foraminifera are close t o zero but several of these coefficients were shown
by Stam (1986) to be significantly greater or less than zero. R- and Q-
mode factor analysis and cluster analysis gave separate assemblages of
mutually associated species. For example, the group with E . mosquensis,
P. rjasanensis, 0 . strurnosum and agglutinants prefers the deep-water
Tojeira shales to the underlying shallow-water Montejunto Formation.
Similar results were obtained by Stam for the Tojeira 2 section.

Additional sampling and Nazli’s autocorrelation analysis

Gradstein and Agterberg (1982) had worked previously with highest

occurrences of Foraminifera in offshore wells drilled on the Labrador Shelf
and Grand Banks. The samples were cuttings obtained during exploratory
drilling by oil companies. Such samples are small, taken over large
intervals and subject t o down-hole contamination so that only highest
occurrences (not lowest occurrences) of Foraminifera can be determined.
These problems associated with exploratory drilling can be avoided on
land if continuous outcrop sampling is possible. According t o
paleogeographic reconstructions (see Stam, 1986), the Lusitanian and
Grand Banks Basins were close to one another during the Jurassic and had
comparable sedimentary, tectonic and faunal history. On land continuous
outcrop sampling can be undertaken in the Lusitanian Basin only.
After preliminary statistical autocorrelation analysis of Stam’s data,
new samples from the two Tojeira sections were collected during the
summer of 1986. F.M. Gradstein identified the foraminifera1 taxa. Only
relatively few samples were taken at exactly the same places where Stam
had sampled before. Figure 3.3 shows typically poor correlations between
proportions estimated from Stam’s and Gradstein’s counts for species in
samples taken at the same spots. These scattergrams reflect random
(binomial) counting errors, local spatial variability of the (unknown)mean
proportion values, as well as possible determination errors. In another
sampling experiment, five samples were taken laterally a t 5m interval
from the same stratigraphic horizon at the base of Tojeira 1. Estimated
 55

ToleIra 1 section Tojeira 2 section

40 1s
70

30 -c 10'
60
50 I '

40
20 I '
30
20
I ..
10
10

0
,::.,..
5 10 15
or:
0
.
10 20 30 40 50 60 70
Eopunulha SPP E mosq~en~i~

40 40

I 70 1
30
6o I
50
20 40
30
10
t
10 20

. . . 10, :..,

5
~

10
~-
15 0 10 20 30 40
0 .0" 10 20
'
30
0
40 0 10 20 30 40 50 60 70
0 SbUmoSUm s Ie""ISElma 0 s,,"m"sl,m s 1e""lSslma

Fig. 3.3 Left side: Proportions of four benthonic Foraminifera for seven replicate samples from same
sites in Tojeira 1 section based on determinations by Stam (horizontal axis) and Gradstein (vertical axis).
Right side: ditto for eleven replicate samples in Tojeira 2 section. See text for discussion of lack of
agreement.

proportion values as well as total benthos counted for these 5 samples were
shown in Agterberg et al. (1990, Table 1). The measured proportions are
markedly different, again illustrating the uncertainty commonly
associated with microfossil abundance data.
As a first step for an M.Sc. project, Nazli (1988) subjected Stam's data
for 14 benthonic species in 31 samples from Tojeira 1 to the ARIMA (Auto
Regressive Integrated Moving Average) procedure of the Statistical
Analysis System (SAS) as implemented on the IBM mainframe computer
at the University of Ottawa in 1986. SAS (Statistical Analysis System) is
a statistical software package with separate versions for mainframes and
personal computers (available from SAS Institute Inc., Box 8000, Cary,
NC, U.S.A.). The ARIMA method was originally developed by Box and
Jenkins (1976). The first part of SAS ARIMA output for E . mosquensis is
shown in Figure 3.4. In autocorrelation, successive values along a time
series are correlated with one another for different lags ( = intervals along
the series). Normally in applications of ARIMA, the values are equally
spaced along the time axis. The decompacted sedimentation rate during
deposition of the Tojeira Formation was about 5cm per 1000years.
Although the shale is homogeneous in composition, it cannot be taken for
granted that sampling it at equal intervals would yield a series with points
56

SAS
ARIMA PROCEDURE
T o j e i r a 1: E. m o s q u e n s i s
AUTOCORRELATIONS

LAG C G V A R I N E CORRELATION
0 160.079 1.00000
1 79.9485 0.49943
2 85.2347 0.53245
3 58.3794 0.36469
4 32.1471 0.20145
5 27.9955 a.174eg
6 14.9058 0.09312
7 25.9934 0.16238
8 23.4033 0.14620
9 19,8307 0.32388
10 12.4919 0.07804

GSC

Fig. 3.4 Partial output of SAS ARIMA procedure for E . mosquensis proportions in Stam's 31 samples
from Tojeira 1 (for complete print-out, see Nazli, 1988, Fig. 4-12, p. 98). ARIMA maximum likelihood
estimation gave three statistically significant coefficients for first order autocorrelation coupled with
two-term moving average. This result is compatible with assumption of signal-plus-noise model in
Figure 3.5.

-0
0.05
1 2 3 4 5 6
, GSC
7
a
lag x

Fig. 3.5 Estimated autocorrelation coefficients of Figure 3.4 plotted along logarithmic scale a n d
approximated by exponential function.

that are equally spaced in time. The 31 samples used for Figure 3.4 are
approximately equally spaced in the stratigraphic direction (see Fig. 3.2,
left side). The resulting autocorrelation pattern for E . mosquensis is
approximately exponential. In Figure 3.4, the first few estimated
autocorrelation coefficients (lags 1 and 2) are greater than zero with a
 57

probability of over 95 percent as indicated by the confidence limits (for two

standard deviations) in the plot on the right-hand side of Figure 3.4. The
approximately exponential nature of the pattern is brought out more
clearly in Figure 3.5 where a logarithmic scale is used for the vertical axis,
so that an exponential function with equation r, = c.exp (-ax)plots as a
straight line. Nazli (1988) has applied other statistical tests including
spectral analysis available a s SAS procedures t o the microfossil
abundance data. He established that most autocorrelation patterns can be
interpreted as white noise (random variability) with the following
exceptions: In Tojeira 1 , E o g u t t u l i n a sp., E . m o s q u e n s i s a n d
O p h t ha1 m id i u m st r u mas u m ex h i b it non-r ando m p a t t e r n s w i t h
approximately exponential autocorrelation functions. E . rnosquensis and
0. strumosum show similar non-random patterns in Tojeira 2 where
exponential patterns were also established for Spirillina tenuissima and
agglutinants. For these seven sequences, straight lines were constructed
on semi-logarithmic plots as exemplified in Figure 3.5 for E . mosquensis in
Tojeira 1. For the three species in Tojeira 1, the analysis was repeated for
a combined series of 41 samples by adding the samples taken in 1986 at
ten new sample sites.
Each straight line was interpreted as representative of a signal-plus-
noise model (cf. Jenkins and Watts, 1968; Agterberg, 1974). The standard
deviation ( S N ) of the noise component for local random variability then can
be estimated from the intercept (c) of the straight line with the vertical
axis. For example, in Figure 3.5, c=0.76. This is the proportion of
variance accounted for by the signal. It leaves a proportion of ( l - c =) 0.24
for the noise component. The variance of the 31 values was 0.0160079 (cf.
Fig. 3.4). Multiplication of this value by 0.24 and taking the square root
yields S N = 6.2 percent. One would expect this standard deviation t o be at
least as large as the standard deviation (sg) arising from the binomial
counting process. The value s g can be estimated from the average
proportion ( = p ) and average number (=ti) of counts per sample. For
example, n =443 for Stam’s 31 Tojeira 1 samples; the corresponding
average proportion value for E . mosquensis is p = 22.5 percent. From the
binomial variance for proportions with equation s 2 g = p (1-p) / n, it then
follows that s g = 1.98 percent. Because for the ratio, sg/sl\r=O.32, this
result would mean that 32 percent of the measured random variability for
E . mosquensis in Tojeira 1 (Stam’s 31 samples only) is due to counting
errors whereas the remaining 68 percent can be ascribed t o local random
variability in the rock. This result is shown in Table 3.1 together with
58
similar statistics for the other species with approximately exponential
autocorrelation functions in the Tojeira sections.

Discussion

Binomial theory h a s been widely used in paleontology and

stratigraphy for estimating the precision of relative abundance with (cf.
Shaw, 1964; Dennison and Hay, 1967). A graph (Fig. 3.1) can be used to
rapidly estimate the probability of not detecting a species if it is present.
Several other graphical methods of calculating sums of binomial
probabilities have been developed. For a summary, see Johnson and Kotz
(1969). The latter publication also contains various approximations for
the binomial, and references t o tables containing values of individual
probabilities and sums of probabilities.

TABLE 3.1

Comparison of standard deviations (in percent) due to counting (sg) and total local random variability
( s ~ for
) species with average proportion jj (in percent) and approximately exponential autocorrelation
function (after Agterberg et al., 1990).

Tojeira 1 (31samples; A=443)

(a) Eoguttulina spp. 2.77 0.76 2.2 0.78 0.36
(b) E.mosquensis 22.47 0.76 6.2 1.98 0.32
(c) 0.strumosum 1.93 0.50 1.7 0.59 0.37
Tojeira 2 (30samples; A = 250)
(a) E . mosquensis 13.84 0.88 3.8 2.19 0.57
(b) S.tenuissima 25.75 0.90 5.5 2.76 0.50
(c) 0.strumosum 11.25 0.91 2.8 2.00 0.71
(d) Agglutinants 10.42 0.58 3.2 1.93 0.61
Tojeira l(41 samples; iL=408)
(a) Eoguttulina spp. 2.20 0.48 2.9 0.71 0.25
(b) E . mosquensis 23.76 0.52 8.4 2.11 0.25
(c) 0.strumosum 2.39 0.60 1.8 0.76 0.41
 59

It should be kept in mind that binomial theory only can provide

approximate estimates of precision of relative abundance estimates. The
main reason for this is that, as when red balls are drawn at random from a
vase with balls of many colors, binomial theory applies t o random
mixtures. In practice, the random variability model only may account for
part of total spatial variability. In this section, a more general model was
applied with X i = S i + N i ; N ~ = N L ~ + N BIt~ is . assumed that at each
sample location (i) an observed proportion value (Xi) is the sum of a signal
( S i ) and a noise (Nil component. The signal is “random” with constant
autocorrelation function as generally is assumed in statistical time-series
analysis and mining geostatistics. (However, a deterministic trend or drift
component also could exist in and might need special consideration). By
systematically comparing relative abundance values for samples taken at
different distances from one another (mainly perpendicular but also
parallel to bedding), it is possible to estimate separate variances of signal
and noise. In the practical example (Tojeira sections, Portugese Oxfordian
black shales), the existence of “signal” could be established for only 2 of 14
species in both sections although 3 other taxa showed systematic change in
abundance through time in one of the sections only. The “noise”
component can be imagined as resulting from local random variability
that arises when samples are taken very close to one another but not
exactly at the same locations. This noise is the sum of the binomial
counting error ( N B ~and ) a local noise component without counting error
( N L ~ )Theoretically
. the latter component is independent of sample size.
In Table 3.1 it is shown that for the 3 taxa with “signal” in Tojeira 1, the
sampling error ( S B ) is about one third of the standard deviation (SN)of
total noise. The ratio S B / S N is close to 0.6 for the 4 taxa with “signal” in
Tojeira 2. Later (in Section 3.6) it will be shown for E . rnosquensis that the
signal can be extracted by eliminating the total noise component.
The purpose of the material presented in this section was not only to
show how binomial theory can be applied t o estimated microfossil
proportion data but also to indicate that probabilities and standard
deviations estimated by means of this theory may be valid only for random
mixtures of microfossils derived from the samples as taken in the field. In
this respect, microfossil abundance data resemble, for example, assay
values in mining for which special geostatistical techniques have been
developed (see e.g. David, 1977).
60

3.4 Multiple pairwise comparison

Hudson and Agterberg (1982) listed several trinomial models by

means of which three probabilities p l , p , and p , (for occurrence of A,, A, or
A,) can be estimated using all possible pairwise comparisons of two
stratigraphic events. Here A, denotes the situation that a n event Ei
occurs above another event Ej in a section, A, is for Ej above Ei, and A, for
the situation that Ei and Ej are coeval. These models include Glenn and
David’s (1960) model, and Davidson’s (1970) model (also see Section 6.10).

Davidson’s model was successfully applied by Edwards and

Beaver (1978) and later by Hudson and Agterberg (1982) t o several data
sets. Drawbacks, pointed out in the latter publication, were that this
method, because of many iterations required, becomes time-consuming
even for digital computers when the number of events exceeds 40. Also,
the model is not able t o handle the situation that many events in the upper
parts of a large stratigraphic column occur with certainty above many
events in its lower parts. Agterberg (1984) showed that a modification of
Glenn and David’s model is not subject to these constraints and can be
used in situations where Davidson’s model is definitely not applicable.

Glenn and David’s model is an extension of the so-called Thurstone-

Mosteller model (cf. Mosteller, 1951) which uses Gaussian curves for the
distribution of positions of events along a linear scale as is done in the
RASC model. The original Thurstone-Mosteller model does not permit
ties. (In stratigraphy ties are coeval events.) As a first step for calculating
average distances between events along this linear scale, the observed
cross-over frequencies are converted t o 2-values according to the
transformation @-‘(P) = 2. This is the inverse of P = @(2)where 0
denotes the fractile (cumulative frequency) of a normal distribution in
standard form. Mosteller (1951) has shown that, under certain conditions,
the best position of an event along the scale is obtained by averaging all
2-values for pairwise comparisons of this event t o all other events. The
resulting position is “best” in a least squares sense. If the RASC model
would be used in a situation that none of the frequencies P,j. are missing or
equal to one, then the unweighted method (simple averaging of 2-values
regardless of sample sizes) would yield results nearly identical t o those of
the Thurstone-Mosteller model. Modifications were made in the RASC
model t o avoid missing values and frequencies equal to one or zero. These
modifications can also be applied t o Glenn and David’s model. This
 61

trinomial model successfully estimated the probability that two events are
coeval in several applications (see Section 6.10).

In the RASC model, observed ties are not ignored but each tie of two
events Ei and Ej is scored as a 50 percent probability that Ei occurs above
Ej and a 50 percent probability that Ej occurs above Ei. Observed scores So
can be compared with estimated frequencies S , = P,x R in which the
estimated probabilities P, (for Ei occurring above Ej) satisfy P, = cP(d,); d,
may be estimated by means of the weighted scaling option of the RASC
computer program in which variations of sample size R are considered.
The agreement between observed and estimated scores was excellent for
Cenozoic Foraminifera on the Labrador Shelf - Grand Banks (see Section
6.10, for details). The chi-squared test for goodness of fit was used for
making this comparison. This shows that the scaling method of RASC
permits the use of significance tests for comparing pairs of events with one
another on the basis of probabilities estimated from the order relationship
of all events considered simultaneously.

3.5 Applications of graph theory

Several authors including Guex (1977), Smith and Fewtrell (1979)
and Agterberg and Nel (198213) have used graphs for representing
relationships between biostratigraphic events . The applications in this
section will be to co-occurrences and superpositional relationships of fossil
taxa. Graph theory is a branch of applied mathematics in which properties
of graphs are established a n d used t o solve specific problems.
Roberts (1976, 1978) has provided an excellent introduction to the topic
(also see Berge, 1973; and CarrB, 1979).

Guex (1987) has made an important contribution to quantitative

stratigraphy by adopting a graph theoretical approach. The Guex
approach differs from the probabilistic one underlying the methods
discussed in this book in that co-occurrencesof fossils are used as the basic
building stones for constructing “Unitary Associations” of fossils which
can be used for correlation. Guex and Davaud (1984, p. 71) stated that
“observed co-occurrences between species must be accepted as true unless
the contrary is demonstrated. No deterministic analysis of the problem
can be performed otherwise”. Later in this volume, results obtained by the
RASC computer program will be compared with results obtained by the
Unitary Associations method for several examples. The purpose of this
62

a b c d e

Fig. 3.6 Example of concepts of graph theory applied in biostratigraphy (after Guex, 1980). (a)
Adjacency matrix containing same information as Fig. 3.6f for sections in Fig. 3.6b; (b) space-time
relationship of 8 species numbered 1 to 8; heavy black vertical lines represent stratigraphic sections with
observations on domains of existence (closed regions) of the eight species; T = time, E = space; (c)
relative chronological position of the intervals I to VI for maximal cliques representing “Unitary
Associations”derived from Figs. 3.6d and 3.6g; (d) matrix relating maximal cliques ( K ) of Fig 3.6g to the
eight species ( X ) ; (el maximal cliques ( K ) identified in four sections (pl-pz) of Fig. 3.6b; (0
biostratigraphical graph G representing co-occurrences and superpositional relationships between the 8
species as observed in the four sections; (g) undirected graph G, representing co-occurrences of Fig. 3.6f
only; (h) directed graph G, with arcs for superpositional relationships. The original purpose of this
diagram was to illustrate, for a simple example, that construction of an interval graph (see Fig. 3.7)
normally does not result in a chronological ordering. Only “reproducible Unitary Associations” are
chronologically ordered as shown in Fig. 3.6e (Guex, 1980).

section is t o introduce the additional concepts of graph theory needed for

this. Figure 3.6 (from Guex, 1980) will be used for illustration.

Graphs consist of vertices and arcs or edges. An arc is an edge with an

arrow indicating the direction for an ordered pair of vertices. Hypothetical
space-time domains of eight fossil species are shown in Figure 3.6.
Observations were made in four stratigraphic sections (heavy black lines
in Fig. 3.6b). All observed relationships of co-occurrence or superposition
are shown in the graph G of Fig. 3.6f which can be decomposed into an
undirected graph (Fig.3.6g, G , with edges only) and a directed graph
(Fig. 3.6h, G, with arcs only). The same information is contained in the
so-called adjacency matrix of Figure 3.6a. Each of the fossils has a row and
 63

a column in Figure3.6a. If two species are observed to co-occur, this is

shown by a pair of ones in the adjacency matrix (e.g. 1 and 2). An ordered
pair (e.g. 4 and 1)is coded by means of a one in the column for 4 (and row
for 1above the diagonal of zeros in Fig. 3.6a) and a zero in the row for 4 and
column for 1 (below the diagonal). If a fossil is observed above another
fossil in one or more sections and below it elsewhere, this pair of fossils will
be scored as a pair of ones in the adjacency matrix.

An undirected graph G, is called complete if it contains all possible

edges. A complete subgraph of a n undirected graph is called a clique. A
clique is maximal if it is not contained in a larger clique. Figure 3.6g has
six maximal cliques labelled I to VI in Figures 3.6~-e. For example, the
subgraph (4,8) is complete in Figure3.6g. It is referred to as maximal
clique VI with two consecutive ones in the matrix of Figure 3.6d. Another
example of a maximal clique is I11 (for fossils 1, 2 and 3) with three
consecutive ones in Figure3.6d. In the example of Figure3.6, the
maximal cliques are “Unitary Associations” which can be recognized in
individual sections without ambiguity (see Fig. 3.6e) and used for

Cmph: Interval assignment:

GI

2-
1 2 4 5

Jfd Jlw/
JfvJ

Fig. 3.7 G1 and Gz are examples of interval assignments A t ) , i = 1, 2, ... for undirected graphs. An
interval assignment for 2 4 with vertices u. u, wand z does not exist (after Roberts, 1976).
64

correlation. In general, the situation is more complex than that shown in

the example of Figure 3.6 and additional concepts and methods of graph
theory are needed.
In general, a set of intervals on the real line can be represented by
means of a so-called interval graph. Only graphs with a interval
assignment (Fig. 3.7 from Roberts, 1976) are interval graphs. The interval
J(i)of a vertex i of an interval graph overlaps a t least in part with the
intervals of vertices to which i is connected by an edge.
The special graph 2 4 (Fig. 3 . 7 ~is
) not a n interval graph because it is
not possible t o assign intervals to it. The vertices of 2, are labelled u, u, w
and 3c in Figure 3 . 7 ~ .According to the preceding definition of a n interval
assignment, the intervals J(u) and J(u) would have t o overlap because u
and u are connected by a n edge. J(u) extends t o the right of J ( u ) in
Figure 3 . 7 ~because it cannot completely lie within J(u) (otherwise, J(w)
could not be overlapping J(u) without overlapping J ( u ) as required).
According to the relationships drawn in Z,, J ( w )overlaps J(u)but not J(u)
and must be depicted in the interval assignment as shown. It is not
possible now t o draw the interval for J(x) which should overlap with J(w)
and J(u) but not J(u). This completes the proof that 2, does not have a n
interval assignment and is not a n interval graph.
A graph Ge with vertices V and edges E can be written as Ge = (V, E ) .
A graph He = (W, F)is a subgraph of Ge = (V, E ) if W is a subset of V and
F a subset of E . He is called a generated subgraph if F consists of all edges
from E joining vertices in W. It can be seen that if G , is a n interval graph,
then every generated subgraph (but not every subgraph) must also be a n
interval graph.
Any graph Ge representing associations of fossil species should be a n
interval graph because pairs of fossils coexisted during specific time
intervals with or without overlap. The question of when a graph is an
interval graph can be answered in several ways. Fulkerson and
Gross (1965) have proved the theorem that a graph Ge is a n interval graph
if and only if there is a ranking of the maximal cliques of Ge which is
consecutive. A ranking K,,K,, ..., K Pof the maximal cliques of Ge is called
consecutive if whenever a vertex u is in K iand Kj for i < j , then for all
i < r < j , u is in K r . It is easy to see that the maximal cliques of Ge in
Figure 3 6 are consecutive. Consequently, Ge of Figure 3.6 is a n interval
graph.
 65

Gilmore and Hoffman (1964)proved the following theorem: A graph

Ge is an interval graph if and only if it satisfies the following conditions:
(a) 2, is not a generated subgraph of Ge, and (b) GeC is transitively
orientable. GeCis the complementary graph of Ge. It has the same vertices
as Ge but edges only between those vertices which are not connected by
edges in Ge. If Ge is a n interval graph, GeChas edges connecting vertices
representing nonoverlapping intervals only. Suppose that arrows are
assigned to these edges thus changing them into arcs either pointing in the
direction for “before” or “after”. It is easy to see that, if Ge is a n interval
graph, these arrows all point either in the forward or in the backward
direction of the real line. Conversely, if GeChas the preceding property,
then Ge (without 2,‘s) is a n interval graph according to the theorem of
Gilmore and Hoffman. The formal definition of a transitively oriented
graph G , is that, if (travelling in the directions of the arrows) a vertex u
can be reached from another vertex u,and a vertex w from u, then w can be
reached from u.
A graph G representing stratigraphic relationships (e.g. Fig. 3.6Q
generally is a mixture of a n undirected graph Ge and a directed graph Ga.
From the preceding two theorems, it can be seen that the complement of
Ge for the example (Fig. 3.6g) is transitively orientable. The directed
graph Ga (Fig. 3.6h) for observed superpositional relationships is a
subgraph of the oriented complement of G,. In a situation that the
relationships between all possible pairs of fossils are fully known, the
biostratigraphic graphG would be the union of G , and its oriented
complement. If Ge is an interval graph, G cannot contain any if a number
of “forbidden” generated subgraphs. For example, the Guex’s cycle C , is a
frequent forbidden structure with 3 vertices (u,u, and w )showing u before
u, u before w and w before u. This is comparable with the 3-event cycle for
stratigraphic events t o be introduced in Chapter 5 on ranking (e.g. cycle
ABC in Fig. 5.7). In a biostratigraphical graphG, C, is not a possible
generated subgraph because it would mean that GeC is not transitively
orientable and Ge is not an interval graph.
C , constitutes the most frequently encountered forbidden structure in
biostratigraphical graphsG. C,’s are likely t o occur in the strong
component of G if it exists. The strong component of a graph is defined as
the generated subgraph which is strongly connected and h a s the
maximum number of vertices. A directed graph is called strongly
connected if for every pair of it vertices u and u, u is reachable from u and u
from u. Guex and Davaud (1984) introduced a special coefficient s = c/r for
66

each arc (e.g. u to u ) where c represents number of times this arc occurs in a
C, within the strong component and r is the total number of times the arc
occurs in the strong component. If the coefficient s of an arc is high, this
may indicate reworking or contamination. If reworking is suspected, u is
omitted in beds where it w a s observed t o occur above u. F o r
contamination, u would be removed from below u.
Guex and Davaud (1984)have developed further rules for interactive
or automated elimination of other forbidden structures from G. For
example, Z, is removed by assuming “virtual” co-occurrence for either a
pair of two or all four of the fossils involved. Two fossil species are said to
co-occur virtually if their co-occurrence was not observed but inferred.
After elimination of all inconsistencies, the biostratigraphic graph G
yields an interval g r a p h G , of which t h e maximal cliques can be
determined. These are the Initial Unitary Associations (1.u.A.’~). They
are called “initial” because Guex and Davaud (1984)added the following
method for combining some of the I.U.A.’s with one another in order to
form the U.A.’s. The I.U.A.’s are identified in sections as previously
illustrated for the Unitary Associations i n Figure 3.6e. A complete I.U.A.
may not be observed i n a section. However a given I.U.A. is fully
characterized by anyone of its unique species or pairs of species. I.U.A.’s
characterized by “virtual’*(inferred, not observed) co-occurrences of fossils
only cannot be identified i n sections. Guex and Davaud (1984)then
proceeded by constructing the directed graph Gk of superpositional
relations between the I.U.A.’s as identified i n t h e sections. T h e
construction of Gk with t h e I.U.A.’s as vertices i s identical to t h e
extraction of Ga for the original biostratigraphical graph G. Next they
find the I.U.A.’s with the longest path in Gk. In general, a vertex in a
directed graph Ga is connected to another vertex by means of a “path” if
the arrows on the arcs between these two vertices point in the same
direction. Each I.U.A. not on the longest path is combined with the I.U.A.
on the path with which it has a n interval in common. This gathering
process yields the final Unitary Associations (U.A.’s) which are identified
in the sections as the I.U.A.’s were before. If the new 1.U.A.-U.A. method
is applied to the example of Figure 3.6, the Initial Unitary Associations I1
and I11 would be combined with one another.
 67

Y Y

Fig. 3.8 Schematic diagrams of cubic interpolation spline and cubic smoothing spline. The cubic
polynomials between successive knots have continuous first and second derivatives at the knots. The
smoothing factor (SF) is zero for interpolation splines. Here as well as in later applications, the abscissae
of the knots coincide with those of the data points.

3.6 Use of cubic smoothing splines for removing "noise" from

microfossil abundance data
Two benthonic species ( E . mosquensis and 0 . strumosum) show
exponential autocorrelations in the Tojeira 1 and 2 sections introduced in
Section 3.3 and are good candidates for attempts to filter out the noise in
order to retain systematic patterns of change of abundance i n the
stratigraphic direction which may be useful for biostratigraphic
correlation. E. mosquensis was selected for further work because it is
relatively abundant throughout the entire shale section of Tojeira 1 and 2
whereas 0. strumosum is nonexistent or rare in the lower half of the
Tojeira Formation.
Various statistical methods are available for elimination of noise
from data. These include curve-fitting using polynomial or Fourier series,
geostatistical "Kriging", signal extraction as in statistical theory of
communication, and the construction of smoothing splines. A variant of
the latter technique will be used here because it is particularly well suited
for coping with the problem of irregular sampling intervals i n one
dimension.
Figure 3.8 illustrates the concepts of interpolation and smoothing
spline functions. Although splines of higher and lower orders can be
constructed, the third-order or cubic spline seems t o be optimum for
68

irregularly spaced sampling intervals (see later). Spline functions have a

long history of use for interpolation; e.g. in numerical integration. Their
use for smoothing is a relatively recent development which commenced in
the late 1960s after the discovery of smoothing splines by Schoenberg
(1964) and Reinsch (1967,1971). Whittaker (1923) had proposed an early
variant.
The interpolation spline curve passes through all ( n )observed values.
Along the curve, there are a number of knots where various derivatives of
the spline function are forced to be continuous. In the example of
Figure3.8, the knots coincide with the data points. A separate cubic
polynomial with 4 coefficients is computed for each interval between
successive data points. These cubics must have continuous first and
second derivatives. After setting the second derivative equal t o zero at the
first and last data points, the continuity constraints yield so many
conditions, that all (4n-4)coefficients can be computed. Smoothing splines
have the same properties as interpolation splines except that they do not
pass through the data points. Instead of this, they deviate from the
observed values by an amount that can be regulated by means of the
smoothing factor (SF) representing the average mean squared deviation.
For each specific value of SF, which can be set i n advance, or
estimated by cross-validation (see Section 10.41, a single smoothing spline
is obtained. In his recent book on spline smoothing and non-parametric
regression, Eubank (1988, e.g., p. 153) discusses that unequally spaced
data points may give poor results for smoothing splines. De Boor (1978)
pointed this out for interpolation splines. In order to avoid poor results
obtained by following cubic smoothing splines to biostratigraphic data for
constructing age-depth curves, Agterberg et al. (1985) proposed the simple
“indirect” method to be discussed in more detail in Section 9.3. The age
data in this approach have relatively large errors while the depths are
irregularly spaced. First, a cubic spline is fitted to the ages using relative
depths (levels) at a regular interval instead of the actual, irregularly
spaced depth measurements. For this purpose the actual depth levels are
equally spaced with interval distance set equal to unity.
A separate spline is fitted to the depth measurements along a depth
scale, but expressing them as a monotonically increasing function of level.
I n practice this second curve is nearly a n interpolation spline.
Combination of the two curves, accompanied by further smoothing if
required, yields the final cubic spline for the age-depth relationship. This
 69

Y
40

30

20

10

-10

-20

-30

-40

-50

-60

-70

-80 I , I I I I 1 I
1 2 x
GSC

Fig. 3.9 De Boor (1978, Fig. 8.1, p. 224) simulated irregular spacing along x-axis by selecting 12 points
(solid circles) from set 49 regularly spaced measurements of a variable (y) as a function of another
variable (x). The optimum fifth order interpolation spline (with 7 knots) provides poor fit except around
the peak.

result is not subject to unrealistic oscillations as may arise in data gaps if a

spline-curve is directly fitted to the data. In the next section, the indirect
method will be applied to microfossil abundance data. These data show
increases as well as decreases in the stratigraphic direction; oscillations
due t o irregular spacing in the stratigraphic direction arise even more
frequently than in age-depth curve applications for which the spline-
curves must be monotonically increasing with age and depth.
The following experiment with interpolation splines illustrates how
the problem of unrealistic oscillations can be avoided, using the indirect
method. It should be kept in mind that the problem of oscillations in data
gaps becomes even more serious if the data are subject to “noise” as in
applications to microfossil abundances. Figure 3.9 is from De Boor (1978,
70

p. 224). In total, 49 observations were available for a property of titanium

(y) as a function of temperature (x). These data points have regular
spacing along the x-axis. Irregular spacing was simulated by De Boor by
selecting n= 12 data points which are closer together on the peak than in
the valleys. De Boor used this example to illustrate that poor results may
be obtained even if use is made of a method of optimal spline interpolation
in which best locations are computed for ( n - k )knots of a k-th order spline.
For the example of Figure 3.9, k = 5 so that 7 knots were used. Although
these seven knots have optimal locations along the x-axis, the result is
obviously poor, because the shape of the relatively narrow peak is reflected
in nonrealistic oscillations in between the more widely spaced data points
in the valleys. De Boor (1978, p. 225) pointed out that using a lower-order
spline would help to obtain a better approximation. In subsequent
applications, use is made of cubic splines only (k=3). Figure 3.10A shows
the cubic interpolation spline for the 12 irregularly spaced points of Figure
3.9 using knots coinciding with data points. Contrary to the 5th order
spline with 7 knots, the new result provides a good approximation.
Deletion of 3 more points from the valleys (Fig. 3.10B) begins to give the
relatively poor cubic interpolation spline of Figure 3.10C which has
unrealistic oscillations in the valleys because all intermediate data points
were deleted.
Figure 3.10 also shows results obtained by applying the indirect
method in the situation that led to the worst cubic-spline result for the
previous example (7 data points, Fig. 3.100. Figure 3.10D is the cubic
interpolation spline for regularly spaced “levels”. Figure 3.10E is a
monotonically increasing cubic smoothing spline with a small positive
value of SF for the relation between x and level. Figure 3.10F is the
combination of the curves of Figures 3.10D and E. The approximation to
the original pattern for 49 values (Fig. 3.9) is only relatively poor in the
valleys where no data were used for control. Unrealistic oscillations were
avoided by the use of the three-step indirect method of Figure 3.10(D-F).

3.7 Biostratigraphic correlation between Tojeira 1 and 2 sections in

central Portugal using E . mosquensis abundance data
Figures 3.11A and B show sequences of samples (combined Stam and
Nazli data) for the Tojeira 1and 2 sections. Distances in the stratigraphic
direction are given i n meters measuring downward from Stam’s
 71

Y Y Y

50-1 A 501 B

,:;if Ji;(
1 , , 1 1 X
0:5 1 1.5 2 215 0 0.5 1 1.5 2 2:5

Y X Y
50

20

10 0 5 ..
0
X
0 2 4 6 8 1 0 0 2 4 6 8 1 0 0 0 , 5 1 1 5 2 2 5
LEVEL LEVEL GSC

Fig. 3.10 Top part Cubic interpolation splines with knots a t data points fitted to irregularly spaced
data. (A) Use of same 12 points as in Fig. 3.9 gives good result; (B) deletion of 3 points in the valleys still
gives fair interpolation spline although local minima at both sides of the peak are not supported by
original data set of 49 measurements; (C) deletion of 2 more points in the valleys results in poor cubic
interpolation spline. Bottom part: Indirect method of cubic spline-fitting. (D)The six intervals along
the x-axis between data points were made equal before calculation of cubic interpolation spline; (E)non-
decreasing cubic spline with small positive value of smoothing factor (SF = 0.038) was fitted to interval
as function of “levels”; (F) curves of (D)and (E)were combined with one another and re-expressed as
cubic spline function which does not show the unrealistic fluctuations of the cubic interpolation spline of
Fig. 3.10C.

stratigraphically highest sample (No. 6.29)in Tojeira 1. This sample was

taken just below the base of the overlying Cabrito Formation. The
stratigraphically highest sample in Tojeira 2 (No. 11.19)occurs about 6m
below this base. It is noted that 3 samples taken by Stam in Tojeira 2
above No. 11.19 (cf. Fig. 3.2,right side) contained too few Foraminifera for
abundance data to be determined.
The data for E . mosquensis plotted in Figure 3.11, were tabulated in
Agterberg et al. (1990,Table 3). As shown by Nazli (19881,Tojeira
microfossil abundances are normalized when the probit transformation is
applied. (The probit transformation consists of converting a proportion to
72

PROBIT ( r F R A C T I L E + 5) PROBIT (=FRACTILE + 5)

0
8.0 6.0 4.0 3.0 2.0 1.0
0 ,
’ L.7
I

-:
e
.I c
.-0
0 20
.-U
? 40

2z
0 80
U
.-
C

1201
E 80
UI I
fn I
6 I
4 100 I
-C
.-0 I
I
g.; os,
0
0 120 I
I
,mu N
I ”
.2 :
140 :. 14c 0 1 0

0
E
.-
0
I-
0 0
O \
\;
Y Y
GSi
180 18C

Tojeira 1 section T o j e i r a 1 and 2 sections

E. m o s q u e n s l s E. m o s q u e n s i s

Fig. 3.11 Left side: Indirect method of cubic spline-fitting illustrated in Fig. 3.10 (D-F) applied to
probits of E . mosquensis abundance data for Tojeira 1 section. Right side: Same with observations and
spline-curve for Tojeira 2 section superimposed. Patterns were slid with respect to one another until a
reasonably good fit was achieved. Zero distance (at sample 6.29 in Tojeira 1) falls just below base of
overlying Cabrito Formation (cf. Fig. 3.2). Correlation between the two sections is poorest along the 35m
data gap in Tojeira 2.

its fractile of the normal distribution in standard form and adding 5 to the
result). The purpose of the latter expression is to reduce the relative
influence of both relatively high and low values. Such “normalization” is
desirable because smoothing splines are fitted by using the method of least
squares in which the influence of each deviation from the curve increases
according to the square of its magnitude. The smoothing factor (SF)
should not be mainly determined by relatively few values only.
Results for the indirect method applied to E . mosquensis in Tojeira 1
and 2 are shown in Figures 3.11A and B, respectively. The two spline-
curves were slid with respect t o one another until a “best” fit was found
(see Fig. 3.11B). A 10m downward movement of the Tojeira 2 sequence,
which places the base of the overlying Cabrito Formation in nearly the
same stratigraphic position in both sections, produces the best correlation.
 73
It is noted that there is a 35m data gap in the Tojeira 2 section so that the
local maximum and minimum located within the equivalent of this gap in
Tojeira 1 could exist in Tojeira 2 as well. For Tojeira 1, sampling was
restricted to the shales of the Tojeira Formation whereas samples for the
underlying Montejunto Formation in which E . mosquensis is absent or
rare were also obtained and used for Tojeira 2. In real distance, the two
sections are about 2km apart. It may be concluded from the pattern of
Figure 3.11B that it is likely that both Tojeira 1 and 2 share essentially
the same relative changes in abundance of E . mosquensis during
deposition of the approximately 70m of late Jurassic shale in this part of
the Lusitanian Basin.
Stam’s (1986) plots for the P/B (planktonhenthos) ratio in the Tojeira
sections suggested that there may exist several oscillations with peaks
where benthos and plankton are nearly equally abundant separated by
valleys with little or no plankton. Precise correlation of these peaks and
valleys is not possible because of “noise” which even became more
prominent when P/B ratios for Nazli’s samples were added. Agterberg et
al. (1989) showed results obtained by the indirect method of spline fitting
applied to the transformed data for P/B ratio in the two sections. Locations
of samples were shown with respect to Stam’s sample 6.29 in both sections
(Tojeira 2 was slid 10m downward as in Fig. 3.11B). Although, on the
average, more plankton was deposited in the area of Tojeira 2, the spline-
curves display patterns that can be interpreted as similar. In total, there
were probably four peaks in the PA3 ratio indicating successive periods of
planktonic bloom during deposition of the upper Jurassic shale. This
result collaborates the one described for the E . mosquensis abundance data
(see Fig. 3.11).
Not only abundance data can be used for correlation. Reyment (1980)
has reviewed basic techniques combining statistics and time series
analysis applied to morphometrics of evolutionary sequences. Ecologically
induced changes in morphology may be useful for biostratigraphic
correlation as well.

3.8 Multivariate methods

Multivariate methods of correlation, using sample by sample
matrices of similarity, or distance coefficients, seek clustering of samples
(Q-mode) as a function of comparative fossil content. In the final
74

dendrogram, the level of clustering of samples may be selected according

to a value which is a function of the degree of association of the original
taxa observed. Biostratigraphic fidelity is a simple numerical expression
of the preference of a species for a particular cluster (zonal) unit.
Depending on the similarity coefficient and weighting procedure selected,
multivariate cluster analysis and -expression of biostratigraphic fidelity
for taxa in the final dendrogram will define assemblage type zonations.
Excellent reviews were given by Hazel (19771, Brower et al. (1978) and
Millendorf et al. (1978). Individual dendrogram clusters may be either of
paleoecologic or stratigraphic significance, or both. The same is true for
multivariate clustering. on species by species matrices (R-mode). The
latter may be insensitive to rare and scattered first and last occurrences of
taxa, but such may be a n advantage for robust correlation. R-mode
clustering may be successfully applied to small data sets. Multivariate
methods have been reviewed by Brower (1985a). For applications to
chemical determinations and borehole logs, see Reyment and Sturesson
(1987).
Methods of multivariate analysis including principal components
analysis, factor analyses, multidimensional scaling, correspondence
analysis and cluster analysis are firmly based on relatively simple
statistical theory (Kendall, 1975b). Computer programs are widely
available for these techniques which are used extensively mainly outside
the earth sciences. Hohn (1978, 1985) used principal components for
stratigraphic correlation. Order of stratigraphic events in time is not
necessarily preserved when multivariate statistical methods are applied.
For example, Brower (1985a) obtained four clusters (A, ByC and D) for a
data set of Upper Cretaceous Foraminifera from the Western Interior
Seaway of the United States. These clusters clearly identify assemblages
of similar fossils but their order in the dendrogram (A, C, B, D) is not
according to their order in relative geological time which is A, B, C, D.
Nevertheless, the clusters are useful for lateral tracing.
Palynologists have developed a method of stratigraphically
constrained cluster analysis which has proved particularly satisfactory for
pollen frequency d a t a (Grim, 1987). A s opposed t o o r d i n a r y ,
unconstrained analysis, only stratigraphically adjacent clusters are
considered for merging. Grim’s (1987) computer program CONISS for
stratigraphically constrained cluster analysis uses the method of
incremental sum of squares. As an option, this program will also perform
an unconstrained analysis which can be useful for comparison because this
 75

option can indicate re-occurrence if a pollen assemblage higher up in the

sequence.
Another recent example of application of multivariate analysis in
biostratigraphy is provided by Bonham-Carter et al. (1986). Foraminifera1
data from 36offshore wells on the Labrador Shelf, Grand Banks, and
Scotian Shelf were analyzed statistically for biostratigraphic correlation
and for systematic trends in distribution related to paleobiogeography.
Ranking and Scaling (RASC) of the data allowed the recognition of
reliable assemblage zones, grouped for this analysis into six well-defined
time slices. Subsequent application of correspondence analysis using
Hill’s (1979) computer program DECORANA (for D E t r e n d e d
CORrespondence ANAlysis) showed clearly geographic trends in faunal
distribution, differing according to latitude. About one-half of the taxa are
planktonic; many of these restricted to southern and more offshore wells
that were influenced by the presence of a proto-Gulf Stream. The
remaining taxa are predominantly benthonic, and may be allocated
broadly to two groups, one with widespread species occurring throughout
the region, and. a smaller group that is restricted to northern wells on the
Labrador Shelf, possibly favored by the influence of terrigenous sediment
supply. This threefold effect of southern planktonics, ubiquitous
benthonics, and minor northern benthonics is recognized throughout the
Cenozoic, with minor fluctuations. During Middle-Late Eocene, relatively
many taxa are restricted northerly benthonics, reflecting the fossiliferous,
thick terrigenous mudstone sequence in northern wells. During Early-
Middle Miocene, the southerly restricted planktonics predominate,
reflecting Gulf Stream influence during climatic warming. In the late
Neogene, a small group of benthonics are relatively ubiquitous due to the
onset of the shelfbound Labrador current. In this study the combined use
of RASC and correspondence analysis provided a good tool for
unscrambling the influence of both time and paleoenvironment on the
dataset.
Burroughs and Brower (1982) applied Wilkinson’s (1974) method of
seriation t o order a data matrix consisting of the presencelabsence of m
taxa taken from n samples in p stratigraphic sections. The objective of
seriation is to arrange the data into a range chart with the taxa in the
columns and t h e samples i n the rows. This is accomplished by
concentrating the presences of the taxa along the main diagonal of the
matrix so that the range zones are minimized. Bonham-Carter et al.
(1986) showed that Wilkinson’s seriation method may give results similar
76

to Hill’s method of correspondence analysis. Brower (198513) has pointed

out that seriation was originally developed by archaeologists who only
rarely possess information on the sequence of the taxa in individual
sections. Burroughs and Brower (1982) found that ordinary seriation
generally yields solutions in which the originally observed relative
stratigraphic position of the samples within the individual sections has
been lost. They proposed a new method of constrained seriation in which
the order relationships of the samples in the sections is preserved in the
final solution. Bonham-Carter et al. (1986) approached the same problem,
by subdividing their events into six separate time slices on the basis of
prior stratigraphic analysis with RASC. The relative position of events
within any particular time slice remains uncertain so that clusters of
events were more appropriate than a complete stratigraphic ordering of
each event in their study.

3.9 Research on time-scales

The construction of good regional and global time-scales provides a
key theme for further research in quantitative chronostratigraphy.
During the last few years of existence of IGCP Project 148, participants
began work along these lines, because it was realized that an ultimate goal
in stratigraphic correlation is isochron contouring.
Time-scale research falls into two categories:
1. Calibration and linkage of biostratigraphic and other unique
geological events to a common chronostratigraphic scale;
2. Stretching of the (relative) chronostratigraphic scale, along the time
axis, t o create a geological time scale measured in Ma (106y) units.
I n t h e absence of d i r e c t r a d i o m e t r i c e s t i m a t e s for m a n y
chronostratigraphic boundaries, geological and statistical techniques have
to be developed t o allow reliable inferences on the numerical age of stage
boundaries. The use of such indirect methods to construct Mesozoic and
Cenozoic scales, applicable both in local basin sequences and in general,
became an important activity in IGCP Project 148.
The relative ordering of events in Earth history is a primary concern
of geologists. On a regional basis, spatial relationships of separate or
overlapping rock volumes are used for accomplishing this goal. The
 77

simplest type of relative time scale is a sequence of ordered events. From

the variable amounts of overlap between rock volumes, or by making
assumptions on rates of sedimentation, it may be possible t o estimate
intervals between events along a relative time axis. For correlation over
large distances between regions or when the rate of change of geological
processes in time is being considered, it is necessary to use the numerical
time scale which is largely based on radiometric ages of variable precision.

In 1982 two time scales were published (Odin 1982; Harland et al.
1982). There is general agreement on the ages along most of these time
scales. The largest discrepancies amount t o about 10 percent of the ages
estimated (also see Section 1.6). Harland et al. (1982) estimated 144 Ma
for the Jurassic-Cretaceous boundary and 590 Ma for the Precambrian-
Cambrian boundary, and Odin (1982) 130 Ma and 530 Ma, respectively.
Such differences are related to the nature of the materials used for dating.

Although they are helpful for pointing out the existence of significant
discrepancies (see e.g. Gradstein et al., 1988), statistical methods cannot
be used t o resolve difficulties related to the nature of the materials used for
dating. Neither can they solve the problem of choosing decay constants in
order to avoid bias in radiometric dating. However, any radiometric
method is subject t o a measurement error which increases with age and is
usually much greater than the uncertainties associated with the relative
ordering of events using methods of stratigraphic correlation (e.g.
biostratigraphic or magnetopolarity methods). The problem of having to
estimate the age of stage and chronozone boundaries from relatively
imprecise isotope determinations remains even if all sources of bias
related to these methods could be eliminated.

Cox and Dalrymple (1967) have developed a statistical approach for

estimating the age of boundaries between polarity chronozones in the
Cenozoic (Brunhes, Matuyama, Gauss and Gilbert Chronozones). A
slightly modified version of their method was used in Harland et al. (1982)
for estimating the ages of boundaries between the stages of t h e
Phanerozoic geological time scale.

This statistical approach is as follows. Suppose that t, represents a n

assumed trial or “estimator” age for the boundary between two stages.
Then the n measured ages t in the vicinity of this boundary can be
classified as ty (younger) or to (older than the assumed stage boundary).
Each age determination tyi or toi has its own standard deviation s i .
78

Because these standard deviations are relatively large, a number (na) of

the age determinations may be inconsistent with respect t o the estimator
te. Only the n, inconsistent ages t,i with t,i < te and tyi > te were used for
estimation by Cox and Dalrymple (1967). These inconsistent ages may be
indicated by letting i go from 1 to n,.
In Harland et al. (1982) a quantity E2 with
n

I=1 (3.6)

was plotted against te in the chronogram for a specific stage boundary.

Such a plot usually has a parabolic form, and the value oft, for which
E2is a minimum was used as the estimated age of the stage boundary.

10

0 5-

00 I I I I I I r X
30 20 10 00 10 20 30 40

00 I I I I 1 I X
-3 0 -2 0 -1 0 00 10 20 30 40

GSC

Fig. 3.12 Weighting functions on basis of which likelihood function can be estimated. A. The function
f c x ) follows from assumption that every age determination is sum of random variables for (1) uniform
distribution of (unknown) true ages, and (2) Gaussian distributions for measurements. B. The function
f&) is for inconsistent ages only. Its log-likelihood function is -E2,
 79

The s t a t i s t i c a l model o r i g i n a l l y proposed by C o x a n d

Dalrymple (1967) may be formulated as follows. Suppose that a stage with
upper age boundary t , and lower boundary t, is sampled a t random. This
yields a population of ages t , < t < t, with uniform frequency density
function h(t). Suppose that every age determination is subject to an error
which is normally distributed with unit variance. In general, the
frequency density function fct) of measurements of which the errors satisfy
the density function for the normal distribution in standard form
satisfies:

(3.7)

Because h(t)is uniform, this becomes

or:

(3.9)

where CP represents the cumulative distribution function of the normal

distribution in standard form. For this derivation, the unit o f t was set
equal to the standard deviation of the errors. Alternatively, the duration
of the stage can be kept constant whereas the standard deviation (0)of the
measurements is changed. Suppose that t2 - tl = 1, then Equation (3.9)
becomes

(3.10)

Graphical representations of A t ) for different values of D were given by Cox

and Dalrymple (1967; Fig. 7, p. 2611). It could be argued that h(x) is not
necessarily uniform and departures from uniformity would affect f ( t ) .
However, one would need very large samples of age determinations before
the choice of a different model for h(x)would be justified.
Suppose now that the true age T, of a single stage boundary is t o be
estimated from a sequence of estimator ages t, by using n measurements of
variable precision on specimens which are known to be either younger or
80

older than the age of this boundary. This problem can be solved if a
weighting function f i x ) is defined. The boundary is assumed to occur a t the
point where x = 0. If one is only interested in the lower boundary of a
stage, Q, { ( t- t,)/o} can be set equal to one yielding the weighting function
f ( x > t , ) = l - @ ( x ) which is graphically shown i n Figure 3.12A.
Alternatively, this weighting function can be derived directly: If all
possible age above the stage boundary have an equal chance of being
represented, then the probability that their measured age assumes a
specific value is proportional t o the integral of the Gaussian density
function for the errors. In terms of the definitions given, any inconsistent
age ty greater than te has x > 0 whereas consistent ages with ty < t, have
x < 0. It is assumed that standardization of a n age tyi or t,i can be
achieved by dividing either (tyi - t,) or (t,i - t,) by its standard error si
yielding xi = (tyi - t,)/s; or xi = (t,i - t,)/si.

Suppose that xiis a realization of a random variable X . The weighting

function f i x ) then can be used t o define the probability Pi= P ( X i = x i ) =
f i x > A x that x will lie in a small interval A x about xi. The method of
maximum likelihood for a sample of n values xi consists of finding the
value of te for which the product of the probabilities Pi is a maximum.
Because Ax can be set equal t o a n arbitrarily small constant, this
maximum occurs when the likelihood function

(3.11)

is a maximum. The so-called log-likelihood function is obtained by taking

the logarithm at both sides of this equation. For the model of Figure
3.12A,

(3.12)

If the log-likelihood function is written as y and its first and second

derivatives with respect to t, as y' and y", respectively, then the maximum
likelihood estimator 2, occurs a t the point where y'= 0 and its variance is
-l/y" (cf. Kendall and Stuart, 1966, p. 43). The log-likelihood function
becomes parabolic in shape when n is large. Su pose that the equation of
!f
this parabola is written as y = a + 6te + c t e . Then the maximum
likelihood estimate t, satisfies t, = -6/2c with variance s2(t,) = -1/2c. It
 81

will be shown by computer simulation experiments t h a t for most

chronograms in Harland et al. (1982) n is sufficiently large and yields good
estimates 0, of the ages of the stage boundaries with corresponding
standard deviations. It can be shown (see Agterberg, 1988) t h a t a
chronogram using E2 represents the maximum likelihood solution for a
filter with equation

(3.13)

where n > te because n, inconsistent ages are used only. This weighting
function is shown in Figure 3.12B. If the corresponding likelihood
function is written as L,, it follows that E2 =-log, L,.
For example, the quantity E2 is plotted in the vertical direction of
Figure 3.13 for the Caerfai-St. David’s boundary example taken from
Harland et al. (1982, Fig. 3.7i). The data on which this chronogram is
based are shown along the top. Values of E2 were calculated at intervals of
4 Ma and a parabola was fitted to the resulting values by using the method

Y Y Y Y
I
I I
4-
I 0 0
I I
00
I
0
rn-s m m+s
I

3-

2-

I
1 -

07
570 580 Ma

Geologic time
GSC

Fig. 3.13 Chronogram for Caerfai-St. David’s boundary example and parabola fitted by method of least
squares. E z = - log-likelihood is plotted in vertical direction. Dates belonging to stages which are older
and younger than boundary are indicated by o and y, respectively. Standard deviation follows from d
representing width of parabola for Ez equal to its minimum value augmented by 2.
82

of least squares. If the log-likelihood function is parabolic, with E2

satisfying

E2 = - a - b t e
-ct2
e (3.14)

it follows that the maximum likelihood estimator is normally distributed

with mean Te = b/2c and variance s2(2,) = 1/2c. It will be shown in the
next paragraph that graphically s(Q might be determined by taking one
fourth of the width of the parabola at the point where E2 exceeds its
minimum value by 2.0 (see Fig. 3.13). The latter result applies t o
parabolas based on La and L. Harland et al. (1982) defined the error of
their estimate by taking one-half the age range for which E2 does not
exceed its minimum value by more than 1.0. This yields a standard
deviation that is ,/2 times as large as the one resulting from La.
A simple proof of the validity of the modified error-range method
illustrated in Figure3.13 is as follows. According t o the theory of
mathematical statistics (Kendall and Stuart, 1961, pp. 43-44), the
likelihood function is asymptotically normal:
1
e y = -exp (-t2/202)
od2n (3.15)

In this expression 9 = L(xlte) and t = te - r;; u represents the standard

deviation of this normal curve centered about r; = 0. Taking the logarithm
at both sides gives the parabola:
2
y = max - 1 /202
(3.16)

where max represents the maximum value of the log-likelihood function.

Setting y = max- 2 gives t = 20. This means that the width of the
parabola at 2 units of y below its maximum value is equal t o 40. The
parabola shown in Figure 3.13 (and subsequent illustrations) is assumed
to provide an approximation of the true log-likelihood function. The
standard deviation obtained from the fitted curve is written as s. In
Figure 3.13, the y-axis has been inverted so that -y = E2points upwards in
order to facilitate comparison with the chronograms in Harland et al.
(1982).

Figure 3.14 shows estimates based on L. The resulting parabola is

almost equal t o the one in Figure 3.13 which was based on La instead of L.
 83

The estimated ages of the Caerfai - St. David’s boundary and their
standard deviations obtained for L , and L also are similar. This
conclusion will be corroborated by a more detailed comparison of the
weighting functions for L and L, a t the end of this section, and by
computer simulation experiments t o be described in the next section.
However, La does not provide a good approximation of L when inconsistent
ages are missing.
A parabolic chronogram is more readily obtained when the consistent
ages are used together with the inconsistent ages as in the method
discussed here. A numerical example of the kinds of differences in results
obtained is as follows. An age estimate based on the chronogram of
Harland et al. (1982, Fig. 3.4h, p. 57) for the Norian-Rhaetian boundary
would be approximately 213 Ma. The corresponding standard error as
reported by Harland et al. (1982) is 9 Ma. The maximum likelihood
method using the same set of 6 data gives an estimated age of 215.5 Ma
with corresponding standard error of 4.2 Ma.

-4

P
0
0
5 -5-
a
-
Y
m
3 -6-

-7-
Y Y Y Y
I
I I I 1
0 0 0 0 0

Fig. 3.14 Caerfai-St. David’s boundary example. Age ( m ) estimated by maximum likelihood method
using L. Standard deviation (s)and width of 95 percent confidence interval are approximated closely by
results shown in Figure 3.13.
84

The chronogram interpreted as a n inverted log-likelihood function

The approach taken is this section differs slightly from the one
originally taken by Cox and Dalrymple (1967) as will be discussed in more
detail now. The basic assumptions t h a t the dates a r e uniformly
distributed through time and subject to measurement errors are made in
both methods of approach. Cox and Dalrymple (1967, see their Fig. 4 on
p. 2608) demonstrated that, under these conditions, the inconsistent dates
for younger rocks have probability of occurrence Ply with:

(3.17)

where erfc denotes complementary error function and T represents true

age of the chronostratigraphic boundary (boundary between geomagnetic
polarity epochs in Cox and Dalrymple’s original paper). The standard
deviation for the measurement errors is written as om. Setting T = 0 and
using the relationship 3 erfc (2/d2)= 1 - CD (2)it follows that:

P (t) = I - @ ( + )
IY
rn
= f(5)
m (3.18)

If t/om is replaced by x , the weighting function shown in Figure 3.12A is

obtained. Consequently, this weighting function can be interpreted as the
probability that an inconsistent age t, is measured for younger rocks.
Likewise, PI,(t) = f(-t/o,) can be defined for older rocks.

Cox and Dalrymple (1967) next introduced the trial boundary age t ,
and defined a measure of dispersion of all inconsistent dates t, with
respect to t , satisfying:

(3.19)

where P d t ) = P$t) if t 2 0 ; and Pz(t) = Pl,(t) if t 1.0. For t, = T , this

quantity is a minimum (see Cox and Dalrymple, 1967, Fig. 5 on p. 2608).
A normalized version of E2 can be directly compared to the theoretical
curve for D2(t, - t,) when the number of inconsistent dates is large. This
normalization consisted of dividing E2 by average number of dates per
unit time interval. It is noted that PI(t) does not represent a probability
density function, because it can be shown that
 85

(3.20)

In this section, E2 is not interpreted as approximately proportional to

D 2 ( t , - te). Instead of this, it is regarded as the inverse of a log-likelihood
function with Gaussian weighting function. For very large samples, good
estimates can be obtained using the inconsistent dates only. For small
samples, however, significantly better results are obtained by using the
consistent dates also and by replacing the Gaussian weighting function by
fi x).

All Gaussian weighting functions provide the same mean age of a

chronostratigraphic boundary when the maximum likelihood method is
used. However, the standard deviation of this mean depends on the choice
of the constant p in exp(-px2). For example, p = 1.0 for fa(x) in Figure
3.12B. Assuming t h a t f ( x ) of Figure 3.12A represents the correct
weighting function, one can ask for which p the Gaussian function
exp(-px2) provides the best approximation t o f i x ) with x 2 0 . Let u
represent the deviation between the two curves, so that

log, {l - @ ( J ) } = - p r 2 +u (3.21)

Minimizing Xu2 for x i = 0.1 h ( k = 1,2...,20) by the method of least

squares gives p = 1.13. Because of the large difference between the two
curves near the origin, p increases when fewer values x i are used. It
decreases when more values are used. Letting k run t o 23 and 24 yields p
equal t o 1.0064 and 0.9740, respectively. These results confirm the
conclusion reached before that a Gaussian weighting function withp = 1.0
provides an excellent approximation to f i x ) .

3.10 Computer simulation experiments o n estimation of the age of

chronostratigraphic boundaries
Computer simulation experiments were performed by Agterberg
(1988) in order to attempt t o answer the following questions: (a) does the
theory of the preceding section remain valid even when the number of
available dates is very small; (b) how do estimates obtained by the method
of fitting a parabola to the log-likelihood function compare to estimates
obtained by the method of scoring which is commonly used by statisticians
86

0 1 2 3 4 5 6 7 8 9 10
1
I
1111
OII 1
I I I
I I I
II
II
Ylll
11l11 1
I Ill I
I
I
1 I I H (a)
1 1 I I Ill1

GSC
Fig. 3.15 Two examples of runs (Runs No. 1 and No. 7) in computer simulation experiment. True dates
(a) were generated first, classified and increased (or decreased) by random amount. Younger and older
ages are shown above and below scale (b), respectively.

(see e.g. Rao, 1973); and (c) how do results derived from the chronograms
in Harland et al. (1982) compare t o those obtained by the maximum
likelihood method.

Figure 3.15 and Table3.2 illustrate the first type of computer

simulation experiment performed. Twenty-five random numbers were
generated on the interval [ O , 101. These numbers with uniform frequency
distribution can be regarded as true dates (T) without measurement errors.
The stage boundary was set equal to 5 ( = mid-point of interval). Values of
T less than 5 belong to the younger stage A, and those greater than 5 t o the
older stage B (see Table 3.2). The measurement error was introduced by
adding t o 'c a normal random number with zero mean and standard
deviation equal to one. As a result of this, each value of T was changed into
a date t . Some values oft ended up outside the interval [O, 101, like 11.197
in the first example (Run No. 1 in Table 3.2 and Fig. 3.15), and were not
used later. In Run No. 1, a single date for the younger state (A) has t > 5 ,
and a date for B has t < 5 . Suppose now, for example, that the trial age of
the stage boundary t, is set equal to 4.6. Then there are 3 inconsistent
ages for Run No. 1 and these are marked by asterisks in Table 3.2. Each
normalized date x = t - t, was converted into a z-value ( = fractile of normal
distribution in standard form) by changing its sign if it belongs t o the
younger stageA. The value of z was transformed into a probability
 87

+
P = @ ( z ) for values of t on the interval [te - 3, t, 31 where @ ( z ) denotes
cumulative frequency of the normal distribution in standard form. The
frequency corresponding t o 3 is equal t o 0.999 of which the natural
logarithm is equal to -0.001. For this reason, values outside the interval
t, +_3yield probabilities which are approximately 1 (or 0 for the log-
likelihood function) and these were not used for further analysis. Thus a
natural window is provided screening out dates that are not in the vicinity
of the age of the chronostratigraphic boundary to be estimated. Most
probabilities are greater than 0.5. Only inconsistent dates (asterisks in
Table 3.2) give probabilities less than 0.5. The value of the log-likelihood

TABLE 3.2

Run 1 for computer simulation experiment. True dates T were classified as younger (A) or older (B) than
true age of stage boundary ( = 5 ) . Dates t with measurement error are compared to trial age ( t , = 4.6).
Inconsistent ages are indicated by asterisks. z = -x for younger rocks (A) and z = x for older rocks (B).
Standard normal z-value is fractile of probability P . Total of logs of P gives value of log-likelihood
function fort, = 4.6.

X
T t ( = t-4.6) 2 P 4, p
4.587 A 4.380 -0.220 0.220 0,5871 -0.5325
7.800 B 8.048 3.448 3.448
2.124 A 2.193 -2.407 2.407 0.9920 -0.0081
0.668 A 2.239 -2.361 2.361 0.9909 -0.0092
6.225 B 5.802 1.202 1.202 0.8853 -0.1218
9.990 B 9.945 5.345 5.345
4.896 A 4.574 -0.026 0.026 0.5102 -0.6730
4.606 A* 6.487 1.887 -1.887 0.0296 -3.5211
0.796 A 0.553 -4.047 4.047
1.855 A 2.526 -2.074 2.074 0.9810 -0.0192
6.292 B 6.923 2.323 2.323 0.9899 -0.0101
3.280 A 1.998 -2.602 2.602 0.9954 -0.0046
2.422 A 1.435 -3.165 3.165
1.397 A 0.912 -3.688 3.688
4.538 A 4.365 -0.235 0,235 0.5928 -0.5230
0.830 A 0.803 -3.797 3.797
6.194 B* 4.033 -0.567 -0.567 0.2854 -1.2540
4.545 A 3.930 -0.670 0.670 0.7490 -0.2890
4.774 A * 4.814 0.214 -0.214 0.4154 -0.8786
0.905 A 0.713 -3.887 3.887
9.763 B 11.197
8.285 B 8.902 4.302 4.302
3.131 A 3.676 -0.924 0.924 0.8224 -0.1955
9.987 B 9.435 4.835 4.835
9.442 B 9.620 5.020 5.020

Total = -8.0397
88

TABLE 3.3

Values of log-likelihood functions estimated for Run 1 and predicted values for parabola fitted by method
of least squares. Initial guesses of extreme values are indicated by asterisks.

TIME LOG-LIKELIHOOD SUM OF SQUARES PREDICTED LLF PREDICTED Ez

(E log P) (EZ)

3.0 -15.58 10.86

3.1 -14.41 9.37
3.2 -13.30 8.00
3.3 -12.27 6.75
3.4 -11.31 5.63
3.5 -16.98 13.54
3.6 -15.83 12.07
3.7 -14.75 10.73
3.8 -13.75 9.52
3.9 -12.81 8.43
4.0 -11.94 7.46
4. I -11.13 6.59
4.2 -10.39 5.84
4.3 -9.72 5.21 5.11
4.4 -9.10 4.69 4.69
4.5 -8.54 4.27 4.32
4.6 -8.04 3.93 -7.98 3.99
4.7 -7.59 3.65 -7.57 3.71
4.8 -7.20 3.44 -7.21 3.47
4.9 -6.87 3.27 -6.89 3.28
5.0 -6.58 3.15 -6.61 3.14
5. I -6.35 3.06 -6.38 3.04
5.2 -6.16 3.02 -6.19 2.99
5.3" -6.02 3.01** -6.05 2.98**
5.4 -5.93 3.05 -5.95 3.01
5.5 -5.88 3.13 -5.89 3.09
5.6* -5.88* 3.24 -5.88* 3.22
5.7 -5.92 3.40 -5.91 3.39
5.8 -6.00 3.59 -5.98 3.61
5.9 -6.13 3.84 -6.10 3.88
6.0 -6.29 4.15 -6.26 4.18
6. I -6.49 4.51 -6.46 4.54
6.2 -6.73 4.94 -6.71 4.94
6.3 -7.01 5.42 -7.00 5.38
6.4 -7.33 5.97 -7.33
6.5 -7.69 6.57 -7.71
6.6 -8.08 7.23 -8.13
6.7 -8.50 7.91
6.8 -8.97 8.65
6.9 -9.47 9.43
7.0 -10.01 10.24

function for te is the sum of the logs of the probabilities as illustrated for
t, = 4.6 in Table 3.2.
Log-likelihood values for Run No. 1 are shown in Table 3.3 with t,
ranging from 3 to7 in steps of 0.1. The largest log-likelihood value is
reached for t, = 5.6 and this value was selected as the first approximation
t,l of the age of the stage boundary. In total, 21 values o f t , with I t, - tel I
< 1.0 were used for fitting a parabola as shown in Figure 3.16. The fitted
-
parabola is more or less independent of number of values used ( = 21) and
width of neighborhood ( =2). However, the neighborhood should not be
made too wide because of random fluctuations (local minima or maxima)
near t, = 3 or 7 (see e.g. Table 3.3). These edge effects should be avoided.
 89

(a)
m-s m mtr
(b) , m;s T m:s , +

H-z
i ; : u
r6 : A

8-
YY Y Y Y
I I
I I I
0 0 0
- 91 2,

GSC
Fig, 3.16 Maximum-likelihood method used for estimating mean of age of stage boundary in Run 1 (data
as in Fig. 3.15). Standard deviation (s) and 95 percent confidence interval also are shown. A. Likelihood
function L was used. B. Chronogram for Run 1 (using La instead of L ) . Note similarity of s and 95
percent confidence interval in Figs. 3.16A and B.

They are due t o the fact that the initial range of simulated time was
arbitrarily set equal t o 10 in the computer simulation experiment. The
peak of this parabola provides the second approximation rn = Ze2 of the
estimated age. The standard deviation ( s ) of the corresponding normal
distribution can be used to estimate the 95 percent confidence interval
rn k 1.96s also shown in Figure 3.16.
The sum of squares E 2 for La, using inconsistent dates only, is also
shown in Table3.3 as a function of t,. The first approximation of its
minimum value is 5.3. The corresponding parabola is shown in
Figure 3.16. The mean age resulting from La is about 0.3 less than the
mean based on L and its standard deviation is nearly the same. It is
fortuitous that the mean based on La is closer t o the population mean ( = 5 )
than that based on L. On the average, the original maximum likelihood
( L )method gives better results (see results for 50 runs given a t the end of
this section).
Younger and older ages generated in each of the first 10 (unit
variance) computer simulation runs are shown in Figure 3.17 together
with their estimated mean and 95 per cent confidence interval using L.
Theoretically, each population mean ( = 5) is contained within the
95percent confidence interval around the sampling mean with a
probability of 95 percent. The means and standard deviations used for
90

Simulated geologic time

0 1 2 3 4 5 6 7 8 9 10
I I I I I I I I I I I

Fig. 3.17 Dates generated in first 10 runs of computer simulation experiment (cf. results for No. 1 and
No.7 shown in Fig. 3.15). Mean and 95 percent confidence interval estimated by maximum-likelihood
method are shown for comparison with true mean ( = 5).

Figure 3.17 are listed in Table 3.4 (Maximum likelihood method with
parabola). Also listed in Table3.4 are the corresponding results for La
(Gaussian weighting function with parabola). The means based on La are
close t o those for L. The estimated standard deviations tend to be either
 91

TABLE3.4

First 10 runs of computer simulation experiment. Comparison of results obtained by fitting parabola and
scoring method, respectively. Standard deviations marked by asterisks are too large (cf. Fig. 3.18B).

Maximum Likelihood Method Gaussian Weighting Function

Run Parabola Scoring Parabola Scoring
No. Mid-point Mean S.D. Mean S.D. Mid-point Mean S.D. Mean S.D.

I 5.6 5.582 0.479 5.554 0.481 5.3 5.269 0.470 5.260 0.500
2 5.7 5.632 0.481 5.663 0.489 6.3 6.190 0.480 6.264 0.500
3 5.1 5.153 0.420 5. I42 0.423 4.8 4.884 0.335 4.828 0.316
4 4.5 4.506 0.W7 4.507 0.452 4.2 4.321 0.395 4.216 0.354
5 5.1 5.070 0.461 5.089 0.466 5.3 5.217 0.482 5.293 0.408
6 4.4 4.419 0.502 4.448 0.505 4.6 4.625 0.749*
7 5.7 5.710 0.531 5.728 0.542 5.8 5.767 3.924*
8 5.2 5.205 0.406 5.200 0.411 5.0 5.025 0.364 5.017 0.408
9 5.0 5.022 0.417 5.018 0.419 5.0 4.966 0.614*
10 4.2 4.231 0.609 4.232 0.623 4.3 4.248 l.OOl*

slightly smaller or much greater. It can be seen from the results for
Run No. 7 shown in Figure 3.18 that the greater standard deviations are
due to a break-down of this particular method of estimation.
R e s u l t s obtained by m e a n s o f t h e method o f s c o r i n g
(see e.g. Rao, 1973, p. 366-374) also are shown in Table 3.4. In our
application of this method, the following procedure was followed. As
before, the log-likelihood was calculated for 0.1 increments in t, and the
largest of these values was used as the initial guess. Suppose that this
value is written a s y . Two other values x and z were calculated
representing log-likelihood values close t o y at small distances and
l o w 4along the t,-axis. The quantities D1 = 0 . 5 ( z - x ) . l o 4 a n d
+
D2 = (x - 2y z). l o 8 were used to obtain a second approximation of the
mean by substracting from the initial guess. The procedure was
repeated until the difference between successive approximations became
negligibly small. Then the standard deviation of the estimate is given by
SD = 1/1021.
For L , the scoring method generally yields estimates of SD which are
slightly greater than those resulting from the parabola method. However,
the difference is negligibly small (Table 3.4). For La, the scoring method
provided an answer in only 6 of the 10 experiments of Table 3.4.
Similar results were obtained for runs in a second type of computer
simulation experiment using variable measurement error (see Agterberg,
1988, for details). In total, 50 runs were made for each of the two types of
92

l j
m-s m m+s
-' I I 1
I Y Y Y

f-
m &

z o +++++++++++++++++++
$
I
0 0 -1
-4
40 4'5 50 55 60 65 70
40 45 50 55 80 65 70
Simulated geologic time Simulated geologic time
GSC
Fig 3.18 Maximum-likelihood method used for estimating mean age of stage boundary in Run 7 (data as
in Fig. 3.15). A. Likelihood function L was used. B. Likelihood function La did not give good result.

experiments. For constant variance of measurement errors, the parabola

method for L gave an overall mean equal to 4.9287 and standard deviation
0.4979 as calculated from 50 means. The corresponding numbers for the
second type of experiment were 4.9442 and 0.5160. The Gaussian
weighting scheme gave overall means equal to 4.9213 and 4.9414 for the
two types of experiments, and corresponding standard deviations equal to
0.5790 and 0.6541, respectively. If the parabola did not provide a good fit
to the function E2, because of zero values around its minimum, the mean
was approximated by the mid-point of the range of zero values in these
calculations. The results of the 50 runs for the two types of experiments
confirm the earlier results described in this section. Additionally, they
show that the Gaussian weighting function (using La) provides results
which are almost as good as the method of maximum likelihood (using L).

3.11 Smoothing of time-scales with the aid of cubic spline functions

When the ages of a number of successive chronostratigraphic
boundaries have been estimated, they can be further improved by
smoothing with the aid of cubic smoothing splines (cf. Section 3.6). The
ages shown in Table 3.5 and Figure 3.19 will be used for example. They
were derived from chronograms in Harland et al. (1982) with the following
relatively minor modifications: (a) if the chronograms for the two
boundaries of a stage are the same, indicating absence of dates for that
stage, the estimate was assigned to a single point mid-way between the
stage boundaries; (b) imprecise estimates for 6 successive Jurassic stages
were not used; (c) when inconsistent dates are missing, the estimated age
was set equal t o the mid-point of the range for missing data in the
 93

TABLE 3.5

Ages and estimated standard deviations used for fitting spline-curve No. 1 shown in Figure 3.19.

Lower boundary of s t a g e Age S.D.

I Maastrichtian (Maa) 72 1.41

2 C a m p a n i a n (Crnp) 84 I . 59
3 Santonian ( S a d 87.5 1.59
4 Coniacian (Con) 88.5 0.88
5 Turonian (Tur) 91 0.88
6 Cenomanian (Cen) 97.5 0.70
7 Albian (Alb) 113 1.41
8 Aptian (Apt) 122 3.18
9 Barremian (Brm)
10 Hauterivian (Hau) I24 2.83
11 Valanginian (Vlg) 1.77
I35
12 Berriasian (Ber)
13 Tithonian (Tth) 145 4.24
I4 Kirnrneridaian (Kim) 151 2.12
15 Oxfordian-(Oxf) I158 5.30
16 Callovian (Clv)
17 Bathonian ( 6 t h )
18 Bajocian (Baj)
19 Aalenian (Aal)
20 Toarcian (Toa)
21 Pliensbachian (Plb)
22 Sinernurian (Sin)
23 Hettangian ( H e t ) 212 4.95
24 R h a e t i a n (Rht) 213 6.36
25 Norian (Nor) 21 8 2.83
26 Carnian ( C r n ) 228 7.78
27 Ladinian (Lad) 238 3.54
28 Anisian (Ans) I242 7.43
29 Scythian (Scy)
30 Tatarian ( T a t ) 246 7.07
31 Kazanian/Ufirnian (Kaz-Ufi)
I253 8.13
32 Kungurian (Kun)
33 Artinskian ( A r t ) 268 4.24
Sakmarian/Asselian (Sak-Ass)

chronogram; and (d) the standard deviation was set proportional to the age
range listed in the summary time scale (Harland et al., 1982, pp. 52-55)
with constant of proportionality equal to 3 d 2.
The fourth modification (d) is based on the earlier considerations
corroborated by the computer simulation experiments proving that the
parabola for La provides an excellent approximation to the parabola for L.
A cubic spline-curve was fitted to the data in Figure 3.19 for the
following reasons. A spline-curve is very smooth because there are no
abrupt changes in the rate of change of its slope; the principle of least
squares is used; and deviations between observed values (crosses in
94

80 100 120 140 160 1 8 0 200 220 240 260 Ma

1-+141

7- Spline-curve 1

819 -
2 82+
10-

I
11112-
4- ~ 1
13 -

ul
a,
14- I I
15116~
u
a
n
0
P
ti l c

23
24 -
25 -
26 ~

27 ~

28/29.

30 -

31132- cretaceous Jurassic

33

Geologic time GSC

Fig, 3.19 Spline-curves fitted to ages of stage boundaries listed in Table 3.5. Spline-curve 1A was fitted
to data for stage boundaries numbered 7 to 27 only.

Fig. 3.19) and spline-curve are permitted to exist but the sum of squares of
these deviations can be regulated; a weight can be assigned to each
observed value. This weight is inversely proportional to the variance of
the observed value.
Let t h e vertical a n d horizontal axes i n Figure 3.19 represent
observations written as x i , yi ( i = 1,..., n ) , respectively. Then t h e
smoothing spline-function to be constructed minimizes

(3.22)
 95

among all functions g(x) under the condition that:

(3.23)

Here the s(yi) are the standard deviations of the values yi. The sum of
standardized deviations S is a random variable approximately distributed
as chi-squared with n degrees of freedom and variance equal to 2n. The
expected value of S, which is equal to n, was used in the applications of this
section.

It can be seen in Figure 3.19 that the fitted spline-curve No. 1tends t o
follow the stage boundaries in the Cretaceous more closely because these
are relatively precise. In places where the uncertainity is great, the
spline-curve tends t o become a straight line. Spline-curve No. 1A shown
also in Figure 3.19 was fitted t o points for stage boundaries between the
Anisian and Cenomanian. It is nearly straight and closely approximates
Spline-curve 1.

Because the intervals between stage boundaries in the vertical

direction of Figure 3.19 are equally spaced, a straight line in this type of
plot would agree with the hypothesis of equal duration of stages.
Harland et al. (1982) applied linear interpolation between relatively
precise stage boundaries (tie-points). The boundaries numbered 1to 7, 27
and 33 were used as tie-points. Because the crosses for boundaries No. 7
and 27 fall slightly to the right of the fitted spline-curves, the estimates

TABLE3.6

Ages used for fitting spline-curve No. 2 based on equal duration of Hallam's ammonite zones in the
Jurassic; without and with tie-points, respectively.

I Stage n. x. Age S.D.

13 Tithonian (Tth) 8 13.4 I45 4.24

14 Kimrneridgian (Kim) 4 14.1 156 0.00
15 Oxfordian (Oxf) 7
6 15' 5'30
16 Callovian (Clv)
17 Bathonian (Bth) 7 17.7
18 Bajocian (Baj) 7 18.9
19 Aalenian (Aal) 3 19.5
20 Toarcian (Toa) 6 20.5
21 Pliensbachian (Plb) 5 21.4
22 Sinernurian (Sin) 6 22.5
23 Hettangian (Het) 3 23.0 208 0.00
96

180 200 220

.. --_
240 260
--
1
2
3 C
4
B
5 I
6
c
7
0
819
5
10

11112 \
I 77h\

13

o) 14 4 24 x
\ J
15/10
P 530'
2 a

al
(51

In \ s
c

\ '\
23
+
' 6 36 1
24
2, 83

yy4
25 + \

26 a
5
27 5

28129

30 7 07+

\ Art
4 24+ Sah ASS
I- - I -~~
80 100 120 140 160 180 200 220 240 260 Ma

Geologic lime GSC

Fig. 3.20 Spline-curve fitted to ages of stage boundaries for Jurassic listed in Table 3.6. This cubic
smoothing spline passes exactly through two tie-points with SD = 0.

obtained by spline-interpolation are younger than those of Harland et al.

(1982) as will also be shown later (see Fig. 3.21).
With respect to the Jurassic time scale, Kent and Gradstein (1985,
1986) have argued that it is more reasonable to assume equal duration of
zones than equal duration of stages. They used Hallam's (1975) ammonite
zones for spacing the stage boundaries in the Jurassic between tie-points
at the base of the Kimmeridgian and Hettangian, respectively. On the
basis of other evidence including data on rates of seafloor spreading in the
Late Jurassic and Early Cretaceous between marine magnetic anomalies
M25 and MO, Kent and Gradstein assumed ages of 156 Ma and 208 Ma for
these two stage boundaries (No. 14 and No. 23), respectively.
 97
The values of xi used for constructing the spline-curve of Figure 3.19
can be modified by using ni for number of ammonite zones per stage
(see Table 3.6). The new values xi shown in Table 3.6 satisfy

xi2 = 12; i = 13, ..., 23

1
130 r I3O

I Spline curve I Spline curve2

lequal stages) lequal zones1 Spl,ne.curve

-i
G'l
ClV
t Ib0

I 0th
- 170

Fig, 3.21 Comparison of spline-curve ages (rounded off to nearest integer Ma values) for Jurassic to ages
estimated by Harland et al. (1982)and by Kent and Gradstein (1985). The asterisks in column 4 denote
key ages of tie-points through which the spline-curve solution was forced to pass. For further
information see Agterberg (1988).
98

where c = 11/62 = 0.1774 represents the ratio of total number of stages

( = 11)and zones ( = 62) in the Jurassic.

The input for spline-curve fitting was further modified by using as tie-
points 156 Ma instead of 151 Ma for the Oxfordian-Kimmeridgian and
208 Ma instead of 212 Ma for the Triassic-Jurassic boundary, respectively,
setting the standard deviations of these ages equal t o zero. As
demonstrated in Agterberg (1988, Appendix 21, the spline-curve has the
property of passing exactly through points of which the standard deviation
is zero. Spline-curve No. 2 with tie-points is shown in Figure 3.20. The
ages of stage boundaries (rounded off t o 1Ma) obtained by three methods
of cubic spline-fitting are shown in Figure 3.21 for comparison with the
other age estimates. Ages for the modified spline-curve (No. 2) for equal
duration of zones but without use of tie-points are shown between those
based on Figures 3.20 and 3.21. The spline-curves all gave 208 Ma for the
age of the Triassic-Jurassic boundary which is younger than estimate of
213Ma in Harland et al. (1982) although the same original age
determinations were used.

The spline-curves yield ages of 138 Ma and 140 Ma for the Jurassic-
Cretaceous boundary which are younger than the 144 Ma age in Harland
et al. (1982) and Kent and Gradstein (1985). This relatively young age is
mainly due to the effect of (a) a relatively young Oxfordian glauconite age
listed as 148.22 Ma in Harland et al. (1982) and a s 145 k 3 Ma in
Armstrong (1978) who, i n t u r n , extracted it from Gyji a n d
McDowell(1970), and (b) 4 other relatively young glauconite ages listed in
Harland et al. (1982) for the Tithonian. If these 5 dates would not be used,
the spline-curves would also give an age of approximately 144 Ma for the
top of the Jurassic. In the beginning of Section 3.9 it was pointed out that
Odin (Editor, 1982) using more glauconite dates estimated a much
younger age (130 Ma) for this boundary. The problem of estimating the
age of the Jurassic-Cretaceous boundary also will be considered in the next
section.

3.12 Statistical significance of ages

The book on a geological time scale by Harland et al. (1982) differs
from earlier publications on the same subject in that it contains tables
with all dates that were used and detailed description of results (e.g.
chronograms) obtained by systematic treatment of the data. In the last
 99

three sections it has been shown that statistical estimation of the ages of
chronostratigraphic boundaries in the geological time scale can be
improved in two ways: (a) the maximum likelihood method can be used for
estimation of the age of individual chronostratigraphic boundaries, and
(b)after estimating the ages of a set of successive boundaries by the
method of maximum likelihood, these can be further improved by using a
cubic spline-curve for smoothing. The resulting methodological
improvements, however, are small in comparison with changes that result
from changing the input data. Harland e t al. (1982) used high-
temperature dates mainly. If low-temperature dates are used (cf. Odin,
Editor, 1982) significantly younger ages are obtained, for some stages,
especially those near the Jurassic-Cretaceous and Proterozoic-Phanerozoic
boundaries.
Haq et al. (1987) provided a new sea level and sedimentary cycles
chart, calibrated t o a new geological time scale for which they used
mixtures of low- and high-temperature dates. This procedure was
criticized by Gradstein et al. (1988) partly because it can be shown that the
low-temperature (glaucony) ages are systematically younger. Odin
(Editor, 1982) had pointed out for one sample (NDS2) that its glauconite
age of 39.6k1.8 Ma is a minimum age and that 1.5 t o 2 Ma should be
added t o it “bearing in mind the long time necessary for the evolution of
the dated glaucony”. Similar corrections may have to be applied to other
glauconite dates as well.
The following statistical experiments performed by the author was
briefly described in Gradstein et al. (1988). In total, 19 low-temperature
and high-temperature dates listed by Harland et al. (1982;Table 3.1, p. 61)
were used to estimate three different ages of the Jurassic-Cretaceous
boundary. The 7 high-temperature dates in this group of 19 dates are
plotted along the top of Figure 3.22, and the 12 low-temperature dates
along the bottom. The maximum likelihood method was applied taking
the high- and low-temperature dates separately, and t o the combined
group of 19 values. Best-fitting parabolas are shown in Figure 3.22. Trial
ages te at intervals of 4 Ma were used. Detailed calculations are shown in
Table 3.7 for t e = 132 Ma for high-temperature dates only. The parabola
fitted to the log-likelihood values of the high-temperature dates shows a
relatively poor fit mainly because these values are determined, to a large
extent, by a single Jurassic date (153.32f 5.00 Ma). The other older date
100

0-

-5 -

U
0
-y"
L
-10-
.-
-I
do
-I
-1s -

Fig, 3.22 Maximum likelihood method used for estimating age of Jurassic-Cretaceous boundary. See
text for further explanation.

(171.66k9.80 Ma) is too far removed from the Jurassic-Cretaceous t o

make a significant difference.
The glaucony dates separately give a mean age of 133.2k2.3 M a
(error is one standard deviation) which is close t o Haq et al.'s (1987)
estimate of 131 Ma for the Jurassic-Cretaceous boundary. The high-
temperature dates give 147.3 & 5.4 Ma which is close t o the estimates of
144 Ma by Harland et a1 (1982) and Kent and Gradstein (1985). The
estimate based on all 19 dates is 136 k 1.8 Ma. It is close to Harland et al.'s
(1982)chronogram age of 135 Ma. Harland et al. rejected this chronogram
age in favor of their 144 Ma age for the Jurassic-Cretaceous boundary
because of the former's relative lack of precision. The 144 Ma estimated
was obtained by linear interpolation between tie-points for the Aptian-
Albian ( = 113 Ma) and the Anisian-Ladinian ( = 238 Ma) boundaries.
The difference between the 133.2k 2.3 Ma low-temperature and the
147.3& 5.4 Ma high-temperature estimates of Figure 3.22 has its own
normal distribution with mean of 14.1 Ma and standard deviation of 5.8
Ma. In the absence of bias, this mean difference would be approximately
zero. Its standardized value (14.1l5.8=2.93) exceeds the 99% confidence
limit (=2.33) of the z-test for testing a difference between two means for
statistical significance. Statistically, it is therefore 99% certain that the
 101

glauconite-based maximum likelihood age is different and younger than

the one based on the high-temperature isotope ages in agreement with
other comparisons reported in Gradstein et al. (1988).
A s pointed out in Section 3.9, Harland e t al. (1982) gave a
quantitative estimate of the error in the age obtained from a chronogram
by taking this error as half the age range for which the error did not exceed
its minimum value by more than 1.0. They pointed out t h a t the
significance of this error is readily seen where only two identical ages
determine a boundary, one of these being from the youngest stage, the
other from the older stage. From Equation (3.6) for computing E2,this
quantity is zero at the boundary and rises t o 1.0 on both sides of the
boundary when the trial age differs from the experimental age by the
quoted error. By using the concept of maximum likelihood it was shown
that the error of Harland et al. is approximately d 2 times larger than the
standard error, provided that the number of dates is sufficiently large so
that the chronogram has become parabolic in shape.
The following slight modification of the preceding argument by
Harland e t al. also results in a modified estimate of the standard
deviation. Two identical ages at a boundary, one from the younger and the
other from the older stage, can be averaged to provide a single estimate of
the age of this boundary. If the standard deviations of the two age
determinations are equal, their average will have a standard deviation

TABLE 3.7

Calculation of logs of probabilities ( P ) for trial age of 132 Ma using 7 high-temperature dates only. The
sum of these values is one of the values plotted in Figure 3.22 and used to fit the parabola for high-
temperature dates. Procedure is similar to the one followed in the example of Table 3.2. However, every
z-value for an age was obtained after dividing the deviation from the trial age by the measurement error
(s) which previously was equal to unity for all deviations in Table 3.2. A and B represent Cretaceous and
Jurassic material, respectively.

A 119.66 4.00 -3.09 0,001 -0,001

A 125.26 6.00 -1.12 0.131 -0.140
A 132.51 12.00 0.04 0 516 -0.726
A 136.50 2.50 1.80 0.964 -3.324
A 130.87 4.35 -0.26 0.397 -0.506
B 153.32 5.00 -4.26 0.000 -0.000
B 171.66 4.80 -8.26 0.000 -0 000
102

which is d 2 times smaller than the errors of the individual ages. This
result is in agreement with the maximum likelihood approximation of L
by La.
Various authors have assigned different meanings t o the error on the
Mesozoic and Paleozoic time scales of Harland et al. (1982). For example,
Carr et al. (1984) assumed that Harland et al. (19821, by stating that this
error is 2.5 Ma, estimated the age of the Jurassic-Cretaceous boundary and
95% confidence interval as 144k2.5 Ma. On the other hand, Menning
(1989) quotes “confidence limits” for this boundary as 1 4 4 k 5 Ma. The
standard error corresponding to the error of 2.5 Ma estimated by Harland
et al. is (2.5/d2=) 1.77 Ma. Multiplication of this standard error by 2
gives a statistically-based estimate of 144 k3.5 Ma for the 95% confidence
interval. This width is between those of Carr et al. (1984) and Menning
(1989), respectively.
In order to estimate the precision of the ages of chronostratigraphic
boundaries, it is important to have good estimates of the errors of the
isotopic dates on which these age estimates are based. Harland et al.
(1982) found that although most determinations quote a n error, a
significant number do not. Errors for these determinations were
estimated by fitting a linear regression line to the available errorhime
data.
For those isotopic ages that have published errors, it may not be
immediately obvious whether these are standard deviations or 95%
confidence limits. For example, Harland et al. (1982) used a number of
Ordivician and Silurian fission track ages from McKerrow et al. (1980)
with quoted errors of about 10 Ma. In Gale et al. (1980), these same ages
are tabulated with errors “at the 20 level” that are twice as large (about 20
Ma). From this, it can be inferred that the age determination errors in
Harland et al. (1982) are indeed standard deviations, although they were
not identified as such in McKerrow et al. (1980).
If errors are standard deviations, it generally can be assumed that
there is 68 percent probability that the unknown true value occurs within
the error interval reported. By taking error limits that are twice as large
this probability is increased to 95 percent. It should be kept in mind that
statements of this type imply that the error distributions are Gaussian or
“normal”.
 103

CHAPTER 4
CODING AND FILE MANAGEMENT OF STRATIGRAPHIC
INFORMATION

4.1 Introduction
During the past five years it has become common practice t o use
microcomputers for the creation, updating and quantitative analysis of
stratigraphic information. Lists of fossils and stratigraphic events
observed in wells or outcrop sections can be coded and stored together with
measurements on their position. The resulting files can be readily
submitted t o various types of data processing. In the Microsoft Disk
Operating System (DOS), for example, files are identified by filenames
which are from one to eight characters long. These filenames may be
followed by extensions consisting of a period followed by one, two or three
characters.
In order to illustrate data management in biostratigraphy, a number
of datasets ranging from small and simple, to large and complex will be
introduced in this chapter. Later, these same datasets will be used t o
illustrate automated stratigraphic correlation techniques. The primary
purpose of the data management required is to create various types of
sequence files for different stratigraphic sections which can later be
systematically compared with one another in preparation of automated
stratigraphic correlation. Before presentation of the datasets, five types of
files are defined which will be used in the examples. For convenience, the
different types of files are indicated by three-letter extensions as in
Microsoft DOS.

4.2 Five basic types of files

The five basic types of files to be distinguished are: DIC, DAT, SEQ,
PAR, and DEP files.
A dictionary file (DIC) is an ordered list of names of taxa or events.
The sequence position numbers of the items in the list provide unique
104
identifiers for coding purposes. Data (DAT) files contain coded
stratigraphic information for taxa using formats which closely reflect
original data collection procedures. Sequence (SEQ) files are lists of
successive or coeval stratigraphic events which can either be coded
directly or derived automatically from DAT files. Parameter (PAR) files
contain the settings of switches and values of parameters required for
running the RASC computer program for RAnking and Scaling or other
data analysis procedures. Depth (DEP) files contain stratigraphic data for
individual wells or sections, augmented by regional time-scale
information for automated stratigraphic correlation.

As input, the RASC computer program requires a DIC file for

stratigraphic events and a SEQ file for their superpositional relations
within individual sections. Although SEQ files can be coded from original
data records, it is usually more convenient to create DAT files instead of
SEQ files, especially if the information is t o be extracted from large
databases. Depth data can be extracted from a DAT file if automatic
stratigraphic correlation between sections is to be performed on the basis
of probable dephts derived by analysis of DEP files.

DIC files
Dictionary (DIC) files contain lists of fossil names (or event names).
They include all names to be used for a regional study. The order of the
names in the DIC files is arbitrary when the file is created. The names
may be initially ordered according to a system selected by the user. For
example, the alphabetic order of taxa can be used, taxa can be grouped
according to families, with alphabetic order within families, or use can be
made of the order in which different taxa are identified in one or more
relatively complete stratigraphic sections for a region.
Microsoft DOS permits rapid alphabetic sorting of names. (It also is
possible to obtain alphabetic lists by means of RASC.) However, most
stratigraphers prefer other types of order for their lists. When a list of
fossil names, alphabetic or otherwise, is available for a region, the names
can be automatically numbered for the DIC files. The assigned sequence
numbers will later be used as codes for the taxa. It is convenient t o enter
only one name per taxon in the original DIC file for a region. In
exploratory drilling, when well cuttings are used to determine highest
occurrences of taxa (and lowest occurrences are not used because of
 105
downhole contamination), the DIC file initially created for taxa, can be
used for the highest occurrences as well. If both highest and lowest
occurrences of taxa are used, it may be necessary t o create a new DIC file
for events from the DIC file for taxa. A simple procedure for this is t o
automatically replace each taxon dictionary number i (i = 1,2,...,n) by two
numbers (2i-1) and (2i). The odd numbers (2i-1) may be used for lowest
occurrences and even numbers (2i) for highest occurrences. In the RASC
computer program for this procedure the same taxon name is used for
highest and lowest occurrences. They are distinguished in the event
dictionary by preceding them with the indicators HI and LO, respectively.

DAT files
Data (DAT) files contain information on all events in all sections to be
used for the study of a region. Different formats can be used. These formats
may emulate data entry procedures of the paleontologist. DAT files consist
of separate lists of samples corresponding to the separate stratigraphic
sections or wells for a region. Examples of formats are as follows: For
exploratory wells, the paleontologist often works with cuttings which
successively become available while proceeding in the stratigraphically
downward direction. For each well, the depth of a sample, e.g. as measured
from sealevel, can be entered , followed by the highest occurrences of all
taxa identified for this sample. For outcrop sections, the paleontologist
usually works in the stratigraphically upward direction. The distances
measured in the stratigraphic direction (perpendicular to bedding) may be
measured for each region from the base of each section upwards.
Consequently, every section has its own scale. The origins of these scales
which are set at the stratigraphically lowest points in the sections usually
do not occur in the same bed. A common procedure of coding t h e
information consists of entering the name of a taxon followed by its lowest
and highest occurrence measured along the scale for the section. This scale
may be in meters or feet, or may be a sequence of numbers representing
beds counted in the stratigraphically upward direction. If beds without
highest or lowest occurrences are skipped in the counting, the numbers
represent so-called “event levels”. DAT files can automatically be changed
into SEQ and preliminary DEP files. The depth files that can be created
from a DEP file are preliminary because information on probable depths of
events in wells (or probable locations of events in outcrop sections) which
106
is needed for automated stratigraphic correlation only can be added after
application of ranking and scaling to the SEQ file.

SEQ files
Sequence (SEQ) files consist of sequences of all stratigraphic events in
all sections t o be used for the study of a region. The events are positioned
according to their relative stratigraphic position, usually proceeding in the
stratigraphically downward direction. Normally, SEQ files a r e
automatically created from DAT files, replacing them by superpositional
or equipositional (coeval) relations. The relative event levels are used for
indicating order in the SEQ files. The information in a SEQ file is
sufficient to ascertain for any pair of events (A, B) in a section whether A
was observed t o occur stratigraphically above or below B, or whether A
and B were observed to be coeval in this section. SEQ files will be used for
ranking and scaling of the events in the region. In the optimum sequence
for a region, each event will obtain a rank above o r below other events. In
the scaled optinum sequence there will be different intervals between
successive events. Zero interval between successive events along the
RASC scale would indicate that the events are coeval on the average for
the study region.

PAR files
Parameter (PAR) files contain the settings of switches and values of
parameters needed t o run the RASC computer program. For example, the
user may decide t o only use events that occur in k, or more sections. The
value of the parameter k, then has to be set in the PAR file. In some
versions of RASC (e.g. micro-RASC, see Chapter lo), the parameters have
default values which can be changed interactively by the user.

DEP files
Depth (DEP) files contain information on the depths (in meters or in
terms of event levels) of stratigraphic events measured i n t h e
stratigraphically downward direction for single sections. This information
is compared t o the average positions of the events expressed either as
 107

ranks or as RASC distances. Ranks and RASC distances are obtained by

ranking and scaling applied to a SEQ file. If the age (in Ma) is known for a
sufficiently large subgroup of the events used for a region, the RASC scale
can be transformed into a numerical time scale. This may facilitate
interpretation and allows isochron contouring (e.g. automated
construction of lines of correlation for multiples of 10 Ma). Then the
estimated age (in Ma) must be entered into the DEP file. For many types of
applications it may seem to be hazardous to convert scaling results t o the
numerical time-scale. It is not necessary t o change RASC scale into a
numerical time scale for automated stratigraphic correlation. Also, even if
this transformation is applied, the automated stratigraphic correlation
between sections actually remains based on the RASC scale because the
same regional time scale transformation is applied t o all sections. The
RASC scale is subjected to local stretching or shrinking t o change it into a
numerical time scale. In general, the same pattern is obtained for the
lines of correlation based on transformed RASC distances (in Ma) or
original RASC distances. For specific stratigraphic events, it does not
matter whether their probable locations in the sections are based on the
RASC scale or on a numerical time scale derived from it.

1 A-Vaca Valley
8-Pacheco Syncline

i C-Tree Plnos
D-Upper Rellr Creek

i E-New ldria
F-Media Ague Creek
G-Upper Canada

j de Sante Anita
H-La8 Crucee
I-Lodo Gulch
I J-Simi Vslley

Fig. 4.1 Locations of sections of the Sullivan database.

108

4.3: Hay example as derived from the Sullivan database: Lower

Tertiary nannoplankton in California
In his original article on probabilistic stratigraphy, Hay (1972) used
stratigraphic information on calcareous nannofossils from sections in the
California Coast Ranges for example (see Fig. 4.1 for locations). These
sections had originally been studied by Sullivan (1964; 1965) and
Bramlette and Sullivan (1961). The distribution of Lower Tertiary
nannoplankton described in the latter three papers also was used by
Davaud and Guex (1978) and Guex (1987) for testing other types of
quantitative stratigraphic correlation techniques. The original paper by
Hay (1972) resulted in extensive discussions (e.g. Edwards, 1978; Harper,
1981) and applications of other techniques t o the Hay example (e.g.
Hudson and Agterberg, 1982). For these reasons, the Hay example will be
used again here.
Hay (1972) restricted his example t o Lower Tertiary nannofossils for
samples shown on Sullivan's (1965) correlation chart augmented by
stratigraphic information on the Lodo Gulch section from Bramlette and
Sullivan (1961). Several of the nannofossil taxa selected for the example
are known to occur in older Paleocene strata in the Media Agua Creek and
Upper Canada de Santa Anita sections (see Sullivan, 1964). Addition of
this other information to the example changes the relative order of the
lowest occurrences in these two sections. In general, care should be taken
to minimize bias due t o lack of sampling older or younger rocks containing
fossils of which the highest and lowest occurrences are recorded for a
section. This source of bias will be discussed on the basis of the Hay
example. It arises only when the time-span for the example has a length
which is comparable t o those of the ranges of the taxa studied. The
problem is almost entirely avoided in datasets which deal with periods,
rather than ages (see later).
Tables 4.1 and 4.2 are DIC files for the Hay dataset and larger
Sullivan dataset originally coded by Davaud and Guex (1978). Hay (1972)
selected for his examples the lowest occurrences of 9 taxa and the highest
occurrence of one taxon (Discoaster tribrachiatus). The DIC file of Table
4.1 can directly be used as a RASC input file. On the other hand, the DIC
file of Table 4.2 is for taxa only and a DIC file should be created from it
before RASC can be used. Agterberg et a1.(1985) automatically replaced
the number (i) of each taxon by a pair of numbers (2i-1) and 2i for its lowest
and highest occurrence, respectively. For example, taxon 89 (Discoaster
 109
TABLE 4 . 1

Dictionary (DIC file) for Hay example. LO and HI represent lowest and highest occurrences of
nannofossils, respectively.

I LO DISC'OASTER I)ISTINC'TlIS
2 LO C'OC'CC~LlTHllSCRIHELLLJM
3 L O DlSC'OASTE R C;ER M A N ICll S
4 1.0 ('O('C'OLITH1JS SOLlTllS
5 LO ('O( '('OLI T H 1J S G A M M AT ION
h L O RHARDOSPHAERA SCABROSA
7 1.0 DISCOASTER MlNlMlJS
8 L O DIS('0ASTER CRllClFORMlS
9 H I DISC'OASTER TRlBRACHlATllS
10 LO DIS('0LITHUS DISTINCTIIS

tribrachiatus) was replaced by event 177 (LO Discoaster tribrachiatus) and

event 178 (HI Discoaster tribrachiatus). Thus, event 9 in Table 4.1
represents the same stratigraphic event as event 178 in the RASC input
DIC file based on Table 4.2.

TABLE 4.2

Fossil name file (preliminary DIC file) for Sullivan database coded by Davaud and Guex (1978) and
Agterberg et al. (1985). A RASC input DIC file was obtained automatically from this file (see text).

I CHIPHRRGRALITHUS CRISTATUS 27 C!S!OilT.iUS f13BRIATUS 51 RHABOQSPHAERA IRUNCAIR 79 OlSCORSTEA BINOQOSUS

? CHIPHRlGRALlTHUS ACANTHODES
? CHIPHRAGRALIIHUS CALAIUS
4 CHIPHPRGMLITHUS QUBIUS
5 CHIPHHR6MCLIIHUS PROTENUS
6 CHIPHPAGMRLITHUS QUADRRTUS
7 COCCOLITHUS BIDENS
8 COCCOLITHUS CRLIfORNICUS
9 ;OCCOL!IHUS EXPRNSUS
10 CJCCOLIIHUS GFRNQIS
II COCCOLITHUS SOLITUS
12 COCCOLITHUS SIAURIQN
l! COPCOLITHUS 616RS
1 4 coccotirncs UELUS
15 COCCOLITHUS CONSUETUS
16 COCCOLITP!S CPPSSUS
I1 COCCOLITlllS CQIBELLUR
I8 COCCI1LITHJS ERINENS
I q CYCLOCOCi3LITHUS EQnfiATlON
C: CICLJCOCCOLIIHUS LURINIS
:I OISCOLITHUS PECTINATUS
:? ; i s c o t I T w PtAnus
2; 3isio:irws P U L ~ H E R
:4 CISCOL!IHUS PULChEROlQES
2: Dl5:3L:T11115 RlnOSuS
? L BISCOLIIHUS D I S I I N C W
:8 QISCOLIIHUS OCELLRTUS 4 RHRBQOSPHAERR INFLRTR 80 QlSC3RSTER OEfLANQREI
?9 DICCOLII.IJS P4NARIUR 55 ZYGOO ISCUS S l6RO IQES 81 OISCORSIER Q E L I C R W
:b QISCOLIIHUS PUNC-QSUS 56 ZYGOQISCUS RQRNAS 82 QISCOASTER QlASiYPUS
Ti Q I S S O L I ~ H U S SCLIOUS 5? ZYGODISCUS HERLVNI B! OISCORSIER QISTINCIUS
:? DIscoL!:IIcs VESCUS 8 ZY6QDlSCUS PLECTOPONS 84 UISCOASTER FALCATUS
31 QISCOLITHUS VEPSUS 50 iYGOLlTHUS CONCINNUG 85 QISCOASTER LOQOENSIS
34 QiSCOLITHUS P E R T U S l S 60 !VGOLIlHUS CRUX 86 DISCOASTER RULTIRAQIAIUS
:5 UISCCLITII3S E X l L i S LI IYGOLITHUS OISIENTUS 'B DISCORSTER NONRRRQIRIUS
3 UiSCOLITHUS DUOCRI'US 6: ZYGQLIIHUS JUNCTUS 88 DISCORSTER STRAONERI
37 ois:otiiws i n c o w i c u u s 61 ZYGRHRBLITHUS SIMPLEX 89 UISCORSTER I R I B R A C H I A W
38 CYCLQLITIIUS ROBUSXS 64 IYGRHABLITHUS BIJUGRIUS 03 DlSClASTER CRUCIFORRIS
19 ELLIPSOLITHUS MCELLUS 65 BARRUQOSPHAERA 816ELQWI 91 DISCOASTER GERRRNICUS
40 ELLIPSOLITHUS UISTICHUS 66 BRRRUDOSPHRERR UISCULA 92 DISCOASTER LENTlCULRRlS
41 HEL ICOSPHREFI SERlLUflUH 67 nicnmiotirnus FLUS 9: QISCORSTER R R R T l N l l
42 HELICOSPHAERA i O D H O I R 68 RICRANTHOLITHUS INRERUAL I S 04 QISCOASTER MINIRUS
4: ?C:HODCLI'YUS !KEN5 00 MICRRNTHOLIIHUS VESPER 9: 31SCOASTER 5EPTEflRAO:::US
44 LOPHlrQOLlTHUS R E N I T O M I S 70 NICRANTHOLITHUS BRSRUENSIS 0h UISCOASIER SUBLODOENSIS
45 -OP4OOOLITHUS llOCHOLOPHORUS 71 NICRANTHOLITHUS CRENULRIUS 9' QISCORSTER HELIRHTHUS
46 RHABUOSFHREPA CPEBRA 72 RICRRNTHOLITHUS AERUALIS 08 DISCORSTER LlllEATUS
47 RHRDDOSPHAERR #lRIONUE 73 CLRIHROLITHUS E L L I P T I C U S 99 OISCOASIER NEDIOSUS
48 FHA9DCSPHREPA PEPLONGR 74 RHOHBORSTER CUSPIS it0 QlSCOPSiER PERPOLITUS
I? RHABOOSPHIERA RUDlS 75 POLYCLADOLIIHUS OPEROSUS 101 DISCOASIERQIQES KUEPPER:
5" RHANJOSPLIRERA SCABPOSR 7h SPHENOLITHUS MQlRNS IO? DISCCRSIEROIQES MEGRSIYPUS
51 RHRBDQSPHRERR SERIFORMIS 17 FRSCICULQLITHUS INVOLUTUS 10; HELIOLITHUS KLEINPELLI
?: RPREQOSPHRERR I E N U I S 7B OISCORSIER BRRBAUIENSIS 104 HEL IOL I THUS RIEDEL I

Figure 4.2 (after Hay, 1972, Fig. 2, p.261) shows stratigraphic

information for the 10 events of Table 4.1 which occur in the nine sections
110

11
STRATIGRAPHIC INFORMATION
A B C D E F G H I 1 2
n n

< <

Fig. 4.2 Hay example. Highest and lowest occurrences of Lower Tertiary nannofossils selected by Hay
(1972) from the Sullivan database. The 10 events are represented by symbols (cf. Fig. 5.1) which
correspond to numbers in Tables 4.1 and 4.3. 6=lowest occurrence of Coccolithus gammation; 0 =lowest
occurrence of Coccolithus cribellum; 0 = lowest occurrence of Coccolithus solitus; V = lowest occurrence of
Discoaster cruciformis; < =lowest occurrence of Discoaster distinctus; n =lowest occurrence of
Discoastergermanicus; U lowest occurrence of Discoaster minimus; w = highest occurrence of Discoaster
tribrachiatus; A = lowest occurrence of Discolithus distinctus; 8 =lowest occurrence of Rhubdosphaera
scabrosa. See Fig. 4.1 for locations of the 9 sections (A-I). The columns on the right represent a subjective
ordering of the events and Hay's original optimum sequence, respectively.

TABLE 4.3

Two SEQ files for Hay example. Minus signs (or hyphens) denote coeval events (cf. Fig. 4.1). The last
entry for a section is followed by -999. Left side: SEQ file for stratigraphically downward direction.
Right side: SEQ file for stratigraphically upward direction.

A A
9 8 7 6 -5 -4 -3 -2 -1-999 1 -2 -3 -4 -5 -6 7 a 9-999
B B
9 10 -6 - 5 -4 - 7 -3 -2-999 2 -3 -7 -4 -5 -6 -10 9-999
C C
9 1 5 2-999 2 5 1 9-999
D D
10 9 8 5 7 1 2-999 2 1 7 5 8 9 10-999
E E
9 6 4 8 7 3 1 5 -2-999 2 -5 1 3 7 8 4 6 9-999
F F
10 9 8 -7 2 5 -4 3 -1-999 1 -3 4 -5 2 7 -8 9 10-999
G G
9 8 -10 5 -2 -1 4 -3 7-999 7 3 -4 1 -2 - 5 10 -8 9-999
H H
4 9 5 -1 -10 7-999 7 10 -1 -5 9 4-999
I I
10 9 6 4 5 1 -3 2-999 2 3 -1 5 4 6 9 10-999

of Figure 4.1.One or more symbols on the same level in a section in Figure

4.2 indicate that the events they represent cannot be separated. Column 1
on the right side is a subjective ranking based on visual inspection of some
of the more complete sections. Column 2 represents Hay's original
optimum sequence. The order of the events in column 2 is based on
 111

pairwise comparison of the events in the nine sections. An event is placed

above other events if it occurs more frequently above than below these
other events in the sections. This is one of several possible methods for
ranking events (see Chapter 5 ) .

(F)
MEDIA AGUA CREEK

Fig. 4.3 Original stratigraphic information for three sections (F-H) of Sullivan database with
stratigraphic correlation based on nannoplankton faunizones according to Sullivan (1965). Table 4.4
contains information on distribution of 9 taxa in samples from Media Agua Creek section.
112

Table 4.3 shows two possible SEQ files for the stratigraphic
information of Figure 4.2.They are for the stratigraphically downward
and upward directions, respectively. For reasons t o be discussed in
Chapter 5 , the RASC computer program may give slightly different results
for the upward and downward directions. It will be instructive to run the
program on both SEQ files of Table 4.3 in order to illustrate the minor
changes brought about by inverting the order. Such minor changes are
usually much smaller than those resulting from altering the dataset by
resetting switches or parameters in the PAR file (see later). Unless stated
otherwise, we will use SEQ files for the stratigraphically downward
direction which is also the direction in which results are printed out in
tables and graphical displays.
The SEQ files of Table 4.3 contain all information represented in
Figure 4.2. Coeval events are shown by hyphens in the SEQ files. The
RASC computer program reads these hyphens as minus signs. There is
one-to-one correspondence between the SEQ files of Table 4.3 and the
graphical representation of Figure 4.2 in t h a t the latter can be
reconstructed from the former and vice versa. No use was made of a DAT
file in order to obtain the SEQ files from Figure 4.2. This stage can be
skipped for the Hay example because the stratigraphic information is of a
simple nature. Normally, the stratigrapher will wish to construct a DAT
file from which the SEQ file is extracted automatically. This procedure
will be illustrated in the next section.

4.4 Partial DAT file for the Hay example

Figure 4.3 shows three of the sections with positions of samples
studied by Sullivan (1964,1965). For example, a partial DAT file will be
created for section F (Media Agua Creek section) only. Table 4.4 contains
the original stratigraphic information for nine of the ten taxa selected by
Hay (see Table 4.1).Only Rhabosphaera scabrosa was not observed in the
Media Agua Creek section. Hay (1971)used Sullivan’s (1965)Eocene
information only, for samples extending up t o 88 feet below the base of
“Tejon” Formation. According to Sullivan (19641,the Paleocene-Eocene
boundary occurs about 111 feet below the base of the “Tejon” Formation.
Table 4.5 shows two partial DAT files (for Section F only) which were
obtained from the information contained in Table 4.4.The first partial
DAT file (Table 4.5A)shows taxon identification numbers followed by
 113

TABLE4.4

Stratigraphic distribution of nine taxa of fossil nannoplanton for individual samples in the Media Agua
Creek area, Kern County, California (according to Sullivan, 1964, Table 3, and Sullivan, 1965, Table 6).
Stratigraphic distance (D)in feet measured upward and downward from base of “Tejon” Formation;
Paleocene-Eocene boundary occurs between 103 and 118 feet. Fossil (F) numbers in first column as in
Table 4.2; A-abundant; C-common; 0-few; x-rare. Single bar indicates stratigraphic events E l to E l 0
used in Table 4.1 and Figure 4.3 (as defined for samples extending up to 88 feet below base of “Tejon”
Formation); relative superpositional relations are changed by using lowest occurrences of four taxa in
Paleocene shown in lower part ofthe table (also see Table 4.5). Level (L) as in Guex (1987, p. 228).

depths in feet of highest and lowest occurrences. The second file (Table
4.5B)has different depths for the lowest occurrences of five taxa because
the data from the Paleocene also were used. P a r t i a l SEQ files
automatically constructed from the data in Table 4.5are shown in the first
two rows of Table 4.6.The first row (Eocene only) duplicates the row for
Section F in Table 4.3 (stratigraphically downward direction). The SEQ
file in the second row is different from the initial result. It is more realistic
because events 1, 2, 5, and 8 already existed before the Eocene. As
mentioned before, continued use will be made of the original Hay example
114

of Figure 4.2 and Table 4.3 for historical reasons. The extended SEQ file
incorporating the Paleocene data shown in Table 4.6 will be employed as
well. Differences between the SEQ files of Tables 4.3and 4.6 are restricted

TABLE4.5

Examples of partial DAT files for Media Agua Creek section of Table 4.4. Distances (in
feet) measured downward from base of“Tejon” Formation. Guex Levels are shown a s L in
bottom row of Table 4.4.

Distances Guex Levels

A. Fossil
Number
LO HI LO HI
83 88 -522 7 15
17 83 2 7 14
91 88 57 7 9
11 86 -1080 7 17
19 86 -522 7 15
94 72 57 9 9
90 72 -514 9 15
89 88 48 7 9
26 34 -522 10 15

B. Part A modified to consider Eocene and

Paleocene

83 146 -522 5 15
17 257 2 2 14
91 88 57 7 9
11 86 -1080 7 17
19 257 -522 2 15
94 72 57 9 9
90 241 -514 2 15
89 257 48 2 9
86 34 -522 10 15
 115

to sections F and G because these are the only sections with additional
data not used by Hay (1972).
Artificial truncation of the observed ranges of some of t h e
nannoplankton taxa may occur when the coding and analysis are
restricted to relatively narrow time intervals, e.g. for one or two ages. Such
artificial truncation effects should be avoided as much as possible in
practice. It is likely that the relatively large number of coeval events a t
the base of sections A and B in Figure 4.2 is in part also due to artificial
truncation. It is noted that Hay (1972)ignored coeval events in his original
method of obtaining an optimum sequence thus counteracting the possible
truncation effect. In the RASC method, coeval events will always be
considered. Although some ranking methods give the same results
whether or not observed coeval events are considered, the scaling methods
make extensive use of coeval events and these should not be ignored. The
truncation drawback of the Hay example will be avoided in most other
datasets to be discussed later.
The lowest and highest occurrences in the DAT and SEQ files for the
Hay example are based on rare occurrences within samples. Sullivan
(1965)adopted the widely used semi-quantitative method of categorizing
abundance (rare, few, common, abundant) in order to improve upon coding
presences and absences only without following the laborious and possibly
counter-productive, route of actually counting large numbers of individual
fossils. His charts normally show uninterrupted sequences for the
“abundant” and “common” categories (A’s and C’s in Table 4.5), whereas
the sequences for the “rare” and “few” categories (x’s and 0’s in Table 4.5)
are interrupted. As pointed out by Hay (1972),the only reasonable
explanation for the gaps in the sequences of x’s and 0’s is that the presence
or absence of a rare taxon is the realization of a random variable (also see
Section 3.3). All taxa were rare when they first and last appeared in a
TABLE4.6

Partial SEQ files in stratigraphically downward direction for Media Agua Creek section as
derived from partial DAT files ofTable 4.5. Event code numbers a s in Table 4.1.

Eocene l(Distances) 10 9 8 -7 2 5 -4 3 -1

EoceneZ(Guexleve1s) 10 9 -8 -7 2 -5 -4 -3 -1
EoceneandPaleocene 1 10 9 7 4 3 1 8 -2 -5
EoceneandPaleocene2 10 9 -7 4 -3 1 8 -2 -5
116

basin. Some taxa (e.g. F 17 in Table 4.4) never became abundant contrary
to others (e.g. F 89 in Table 4.4) which were abundant as well as rare.
Stratigraphic events can be defined on the basis of rare occurrences as
well as abundant occurrences of a taxon. For example, Doeven et al. (1982)
applied ranking to a mixture of events in order to construct a nannofossil
range chart for Cretaceous nannofossils along the Canadian Atlantic
margin. This mixture included subtops (last consistent occurrences) and
superbottoms (fist consistent occurrences) as well as the tops (last observed
occurrences) and bottoms (first observed occurrences) for selected
nannofossils. Definition of more than two events for these taxa helped to
improve the range chart. In general, subtops and superbottoms are less
subject t o random variability in time than first and last occurrences (also
see Doeven, 1983).

4.5 DAT files constructed by Guex and Davaud

As mentioned in Section 4.3,Guex and Davaud have used Sullivan’s
database for the testing of other types of quantitative stratigraphic
correlation techniques. Their “Unitary Associations” method aims t o
emulate the Oppel zones of biostratigraphy. Oppel (1856) had proposed
construction of a regional standard consisting of a succession of different
zones later called “Oppel zones”. Each zone of this type is characterized by
one or more taxa, or by a unique assemblage of taxa (also see Fig. 2.1 and
previous discussion in Section 2.2). Identification of individual Oppel
zones in individual sections provides a vehicle for biostratigraphic
correlation. As explained in Section 3.5, Guex (1987)used graph theory t o
construct Unitary Associations which have essentially the same properties
as Oppel zones. Systematic insertion of supposedly missing data in order
to establish coexistence of taxa is a guiding principle of this approach. This
aim is already reflected in the type of coding stratigraphic information
performed before the Unitary Associations are constructed.
It is reasonable to assume that, apart from disturbances such as
reworking, each taxon existed continually between the time equivalent of
its observed first and last occurrences in a section. This is the well-known
“range-through” method (cf. Section 2.1) which usually leads to assumed
coexistences of taxa which may not have been observed together within a
single bed. The range-through assumption is made in explicit or implicit
form in most quantitative stratigraphic correlation techniques including
 117

RASC and the Unitary Associations method. However, in the latter

method, the following, additional assumption is made before the data are
coded. Adjoining samples are combined into levels representing “maximal
horizons” (cf. Guex, 1987, p. 20; also see Guex, 1988) as illustrated for the
Media Agua Creek example in the bottom row of Table 4.4.

Davaud and Guex (1987, p. 587) estimated that the number of

“maximal horizons” is less than 30 percent of the total number of samples
for the Sullivan-Bramlette database. Figure 4.4 illustrates how this type
of level was constructed. Each maximal horizon corresponds t o a separate
clique in the interval graph (cf. Section 3.5) for the section that is being
studied. The observed range chart for the section is interpreted as the
interval assignment for this interval graph. The seven taxa in the example
of Figure 4.4 have only three maximal horizons corresponding t o the
cliques (1, 2, 3), (2, 3, 4) and (3, 4, 5, 6, 7) respectively. These maximal
horizons are separated by horizons with fewer taxa on the range chart for
the section.

Individual samples can be represented by lines drawn perpendicular

to the ranges. In Figure 4.4 the taxa whose ranges are intersected by such
a line would coexist in the corresponding sample. All samples containing
taxa of a particular clique are combined with one another as a first step
towards constructing the Unitary Associations. If sampling proceeds in the
stratigraphically upward direction, a new combination of taxa leading t o a
new maximal horizon is started as soon as one or more taxa of the next
clique are encountered in a sample.

An interval assignment of an interval graph is schematic in that

there is no one-to-one correspondence between these two models. In
general, it is not possible to reconstruct the range chart for a section from
its interval graph. For example, when moving from the right to the left in
the range chart of Figure 4.4, one successively encounters 6 , 3 , 7,5, and 4
for the end points of the five taxa in the largest clique. Such detailed
information obviously does not exist in the interval graph.

The eighteen levels “L” in Table 4.5 were based on maximal horizons
for all ( = 82) taxa occurring in the Media Agua Creek area. The 44
samples of this section were combined into 18 levels by Guex (1987) with
loss of information on the relative order of first and last occurrences.
Many pairs of events were made coeval during the coding, although they
had a distinct order in the section before the cliques were determined. For
118

Pig. 4.4 Example of interval assignment J ( i ) , i = 1, 2, ... for undirected graph (after Roberts, 1976). If
applied to a single stratigraphic section, each clique represents a maximal horizon or Guex level.

ranking and scaling generally, it is recommended that all observed

superpositional relations for pairs of events in sections are preserved by
entering this type of information in the DAT files from which SEQ files
will be derived automatically. Table 4.6 shows a partial SEQ file for the
Media Agua Creek section of the Hay example based on Guex levels (line
2) in comparison with that based on all samples (line 1). The number of
hyphens for coeval events is increased when event levels are combined
with one another using the maximal horizons method. For Eocene
nannoplankton only, the number of event levels would be reduced from 6
to 3 in Table 4.6, and for the Paleogene (combined Eocene and Paleocene)
from 7 to 5. Later Guex (1987) added the information for the Paleocene to
the Sullivan data base for the (Media Agua Creek and Upper Canada de
Santa Anita sections. Lines 1 and 2 for Eocene and Paleocene in Table 4.6
show the effect of this change with respect to lines 1 and 2 for the Eocene
used in the original Hay example.
It is noted that Agterberg et al. (1985) made use of the Sullivan
database as originally coded by Davaud and Guex (1978)which did not use
Sullivan’s (1964)data for the Paleocene, and in which the number of levels
had been reduced by adoption of the maximal horizons method.

4.6 Gradstein - Thomas database: Cenozoic Foraminifera in

Canadian Atlantic Margin wells
The RASC model for ranking and scaling of stratigraphic events was
originally developed during a project on Cenozoic foraminifera1
stratigraphy of the northwestern Atlantic margin (Gradstein and
 119

64' 56" 48'

\
t +

I Karlsefni H-13
2 Snorri J - 9 0
3 Herlolf M-92
4 Blarni H-81
5 Gudrid H-55
6 Corlier D - 7 9
7 LeifE-38
8 Leif M-48
9 Indian Harbour M-52
10 Freydis 8 - 8 7
11 Bonavisto C - 9 9
12 Cumberland 8 - 5 5
13 Dominion 0 - 2 3
14 Egrel K - 3 6
15 E g r e t # - 46
16 Osprey H - 8 4
17 Heron H - 7 3
+ I6 Bran1 P-87
19 Kittiwake P - l l
20 Wenonoh J - 7 5
21 Triumph P - 5 0
22 Mohican 1-100

J3

'4.
I5

.I6
+

I
64'
I
56'
I
48.

Fig, 4.5 Location of 22 wells along Eastern Canadian margin used for Cenozoic foraminifera]
stratigraphy by Gradstein and Agterberg (1982). Original samples were obtained from Eastcan and
others: Karlsefni H-13 (1760-12 990'), Snorri J-90 (1260-9950'), Herjolf M-92 (3030-78001, Bjarni H-81
(2760-6060'), Gudrid H-55 (1660-8580'1, Cartier D-79 (1950-6070'); Tenneco and others: Leif E-38 (1210-
3557'); Eastcan and others: Leif M-48 (1300-5620'); BP Columbia and others: Indian Harbour M-52
(1740-10 480'); Eastcan and others: Freydis B-87 (1000-5260'); BP Columbia and others: Bonavista C-99
(1860.11 940'); Mobil Gulf Cumberland B-55 (920-11 830'), Dominion 0-23 (1380-10 260'); Amoco Imp
Skelly: Egret N-64 (1060-2070'), Egret K-36 (860-2270'), Osprey H-84 (1190-2660?, Brant P-8 (1050-
6270'); Amoco Imp: Heron H-73 (970-5800'), Kittiwake P-11 (970.55603; PetroCanada Shell: Wenonah
5-75 (1000-4750'); Shell: Triumph P-50 (990-5490'). Mohican 1-100 (1276-5320').
120

Agterberg, 1982). Figure 4.5 shows the locations of the 22 offshore wells
used. They were divided into two groups. Sixteen of these wells are located
on the Labrador Shelf and northwestern Grand Banks (northern region).
Six occur on the Scotian Shelf and southern Grand Banks (southern
region). In total, the highest occurrences (exits) of 206 benthonic and
planktonic Foraminifera, were used. Of these 150 and 157 occurred in the
northern and southern regions, respectively.

Initial biozonations for the northern and southern regions were based
on smaller sets of 41 and 60 data, respectively. The two regions had 14 of
these taxa in common. The southern biozonation had 32, mostly Eocene
and Miocene index planktonics and the northern zonation 6, essentially
Eocene ones. This difference reflects pronounced post-Middle Eocene
latitudinal water mass heterogeneity and differential post-Eocene
shallowing across the continental margin. The biozonation with relatively
many planktcnics for the southern region helped to establish the initially
largely unknown biozonation for the northern region.

Later, data for 10 wells were added for the northern region, mainly in
the vicinity of the Hibernia oil field on the Grand Banks between wells 13
and 14 in Figure 4.5. New taxa were identified and the original dictionary
for the 22 wells of Figure 4.5 was updated. The enlarged dictionary is
given in Table 4.7 which is part of the Gradstein-Thomas database for 24
wells on the Labrador Shelf and Grand Banks, published in Gradstein et
al. (1985, pp. 515-520).

It is noted that not all events in Table 4.7 are highest occurrences of
Foraminifera. For example, four seismic events were included in the
database. Also, in total there are 238 events in Table 4.7 which is less than
the greatest number (=275) assigned t o a taxon. Gaps in the numbering
are due t o revisions made in the identification of taxa. For example, a
taxon with one name in Table 4.7 may be the composite of two taxa of
which one had a different name which became obsolete after the renaming.
In order t o preserve the unique identifier of the name that was retained, a
dummy code (e.g. xxx) was assigned in the dictionary to the name that was
deleted. The advantage of this procedure is that other taxa retain their
original dictionary numbers in RASC input and output files regardless of
revisions applied t o relatively few taxa.

Table 4.8 is a partial DAT file using 4 of the 24 wells. The depths of
the samples were measured in feet for earlier wells and in meters for wells
 121

TABLE4.7

DIC file of Cenozoic Foraminifera in Gradstein-Thomas database for Canadian Atlantic margin.

1 NEOGLOBOQUADRINR PACHVDERRA 58 EPONIDES spa

2 GLOBIGERINA APERTURA 59 RZEHAK I NA EP I 6 0 N A
J GLOBIGERINA PSEUDOBESR 60 PLANOROTALITES COMPRESSUS
4 GLOBOROTALIA INFLRTA 61 SUBBOTINR PSEODOBULLOIDES
5 GLOBOROTRLIA CRASSAFORlllS h2 GAVELINELLA DANlCA
6 NEOGLOBOQUADRINA ACOSTAENSIS h3 NODOSRRIA S P I I
7 6LOBI6ERlNOIOES RUBER h4 CASSIDULINA ISLANDICA
8 ORBULINA UNIVERSA 65 COSCINODISCUS SP1
9 FURSENrOlNA GRACILIS hh COLEITES RETICULOSUS
10 UV IGER I N 4 CRNAR I ENS1 S 67 SCAPHOPOO S P I
I1 NONIONELLA PIZARRENSE 6E SPIROPLECTAIININA SPECTABLIS LO
12 EHRENBERGINP SERRAIA P9 NOOOSARIA SPB
13 HANZAYAIA CONCENTRICA 70 ALABAIIINA YOLTERSTORFFI
14 TEXTULARIA RCGLUTINRNS 71 EP I STOH I NA ELEGANS
15 GLOBIGERINA PRAEBULLOIDES 72 CVCLOGYRA SPJ
16 CERATOBULIMINR CONTRARIA 73 EPONlDES SP3
17 ASTERIGERINA GURICHI 7 4 EPOhlDES SPS
18 SP IROPLECTAMH I HA CAR lNATA 75 LENTICULINA ULATISENSIS
19 6LOB16ERINOIDES 5 P 75 CASSIDULINA SP
?O GYRO ID I NA 6 I RARDAWA 77 ELPHIOIUfl SP
21 GUITULINA PROBLEM 78 W[GEHINA PEREGRINA
?? COSCINODISCUS SP; 79 GLOBIGERINA TRIPARTITR
23 COSCINODISCUS SP4 80 CYCLARMINI CrlNCELLATl
24 TURRILINA ALSATlCA 61 GLOBIGERINA VENEZUELANA
25 COARSE ARENACEOUS SPP. 82 GLOBIGERINA LINAPERTA
2h UVIGERINA DUIIBLEI 8: PLANOROTALITES PSEUDOSCITULUS
27 EPONlDES UlBONATUS 84 GLOBIGERINA VEGUAENS!S
28 C I B I C I DO I DES SP5 85 PSEUDOHASTIGERINR NICRA
29 CVCLAMMINA RMPLECTENS 86 TURH: L INA BREVISPIRA
20 CIBlC I DO IDES BLANFIEDI 67 BULININA AFF. JACKSONENSIS
31 PTEROPOD S P I 88 SIPIIOGENEEOIDES ELEGANTA
32 AMMOSPHAEROIUINA SPI 89 NOROIOVELLA SPINULOSP
33 TURBOROTALIR POMEROLI 90 RCARlNlNA DENSA
31 M R G I N U L I N A DECORATA 91 R~JIOl&RI&NS
35 SPIROPLECTAMHINA OENTRTR 9? MOROZOVELLA CbUCASlCA
3b PSEUDOHASTI6ERINA YILCOXENSIS 9; ACARlNlNA AFF. BROEDERNRNNI
37 ACARlNlNb RFF PENTACAMERATA 94 GLOBIGERINATHEKA t U 6 L E R I
a: LENTICUL INA SUBPAPILLOSR 95 ARAGONIlr VELASCOENSIS
39 ALABRMINA WILCOXENSIS 96 ACARININA INTERIIEDIR WlLCOXENSlS
40 BULIMINR RLAZANENSIS 100 GLOBIGERINA RIVEROA
41 PLECTOFRONDICULARIA SP1 I09 CASSIDULlNb CURVATA
42 CIB!CIDDIDES ALLEN1 110 GLOBIGEHINA BULLOIDES
43 BUL I H I N R MIDWRYENS IS Ill PARAROTALIP SFI
44 CIB!C!COIDES AFF WEST1 1 I ? IIARGINULINA BACHEI
45 BULIMINR TRIGONALIS 11; GLOBOROTALIA flENARD! I GROUP
46 REGASPORE S P I 114 6LOBI6ERIN010ES SACCULlFkR
47 PLANOROTALITES PLANOCONICUS 11; GLOBOROTAL A I OBESA
4a ANOMLINA SP5 I l b OPBULINA SUTURALIS
19 OSANGULRRIA EXPANSA 117 SPHAEROlDlNA BULLOIDES
50 SUBBOTINA PATAGONICA 118 EPISTOMINR SP5
51 ACARININA P R l M I T l V A 119 SPHAEROIDIWELLA SUBDEHISCENS
52 ACdR I NINA SOL DADOENS IS 120 GLOBOROTALIR SIRKENSIS
53 UVIGERINA BRTJESI 121 6LOBIGER1NA NEPENTHES
:I SPIROPLECTAIIRINA NAVARRORNA I22 SPHPEROIDINELLOPSIS S E l l N U L l W A
55 GAVELINELLA BECCRRIIFORMIS I23 GLOBIfiERINOIDES TRILOBUS
56 GLOMOSPIRA CORONA 124 GLOBORUADRIW DEHISCENS
57 SPIROPLECTAMIIINA SPECTLBILIS L.co 125 m ~ ~CaNiINuosn
~ ~ ~ ~ ~ ~ n
122

TABLE 4.7 (continued)

I26 GLOBIGERINOIDES OBLIRUUS 211 HRNTKENINA SP

I27 GLOBIGERINITA NAPARIMAENSIS 213 ARENOBULIMINA SP?
I28 GLOBOROTAL I R PRAEMENARDI I 216 GLOB1 6ERI NOIDES SICANUS
I30 SIPHONINA ADVENA 217 GLOBOROTALIA SCITULA
131 C l E l C I D O I D E S TENELLUS 218 MARGINULINA AMERICANA
132 'GLOBOROTRLIA' OPIMA NANA 219 MARTINOTIELLA COMMUNIS
133 LENTICULINA SP3 220 C l B I C I D O l D E S HUELLERSTORFFI
134 LENTICULINA SP4 221 GLOBIGERINOIDES SUBWADRATUS
135 6LOBlGERINA SP40 222 GLOBOPUADRINA ALTISPIRA
I36 MELONIS BRRLEANUM 223 GLOBIGERINA CIPEROENSIS
137 GLOBIGERINOIDES PRIHORDIUS 224 UV IGERINR ME X ICANA
138 GLOBIGERINA RNGUSTIUMBILICATR 225 GLOBIGERINA AFF. AMPLIAPERTURA
139 'GLOBOROTALIA' OPIMA OPIMA 2% GLOBIGERINA SENNI
140 ROTALIATINA BULlMlNOIDES 227 C I81CIDOl DES RFF. TUXPANENS IS
141 PLANULINA RENZI 228 CASSIDULINA TERETIS
I42 GYROIOINA SOLDAN11 MAMILLIGERA 230 BULIHINR OVRTR
143 UVIGERINA GALLOYAY 231 UVIGERINA RUSTICA
I44 GLOEOROTALIR CERROAZULENSIS 252 GLOB IGER I N 0 1OES I MMATURUS
145 ANOMALINOIDES ALLEN1 233 CATAPSVDRAX UNICAVUS
I46 SUBEOTINA EOCRENA 234 TRUNCAROTALOIDES RFF. ROHRI
147 CRTRPSYDRRX RFF. D I S S I H I L I S 235 SUBBOT I NA BOL I VRRI ANA
148 GLOEIGERINATHEKA INDEX 236 EPONIOES SP4
149 GLOBIGERINATHEIP TROPICALIS 237 LENTICULINA SPE
I50 GLOBIGERINA GORTANII 238 C I81 C ID0 IDES SP7
I51 BULIMINR BRRDEUPVI 2 3 NONIONELLA LABRADORICA
15: BUL I M I NA COOPERENS IS 240 ELPHIOIUM CLRVATUM
154 ANOMALINOIDES MIDHAYENSIS 241 GLOBOROTALIA TRtiNCRlULINOIDES
155 AN0MALINOlDES GROSSERUGOSA 242 GLOBOROTALIA FOHSl GROUP
1% SUBBOTINR FRONTOSA 243 GLOBIGERINR DECAPERTA
157 TRlTAXlA SP3 244 GBUDRYINA S P l O
158 SUBBOTINA !NAEQUISPIRA 245 PRAEORUULINA GLOMEROSA
159 MOROZOVELLA ARAGONENSIS 24h GLOBIGERINATELLA INSUETR
I60 ACARININA PSEUDOTOPILENSIS 247 GLOB16ERINOIDES ALTIAPERTURA
161 PLANOROTALITES AUSTRALIFORMIS 248 'GLOEOROTRLIA' AFF. INCREBESCENS
lb? I(OROZ0VELLA AEQUA 249 GLOBIMRINATHEKR SEMIINVOLUTR
I h 4 NUTTAL IDES TRUMPVI ?50 VULVULlNd J A R V I S I
!h6 MOROZOVELLA SUBBOTINAE 25 I ANOMALINA SP4
167 MOROZOVELLA FORMOSA GRACILIS 252 MOROZOVELLA AFF. QUETRA
1h9 EPISTOMlNELLA TRKRYANAGI 1 25: SUBBOTINA TRILOCULINOIDES
172 PSEUDOHRSTIGERIIR SP 254 PLANOROTAL l l E S PSEUDOllENARDI 1
I73 ANOMALINA S P I 255 MOROZOVELLA CONICOTRUNCATA
I75 ALLOGROMIA SP 2% 'MOROZOVELLA" AFF. PtiSILLA
176 ALLOMORPHINA S P I 257 CHILOGUEMBELI N A SP
177 B O L l V l N b DILATATA Z5E TAPPANINA SELMENSIS
179 GLOBOROTRLIR SCITULR PRRESCIlUtA 259 AflMODISCUS LRTUS
I a0 GVROIDINA SP4 260 HAPLOPHRAGMOIDES K I R K 1
lEl CYCLOGVRA INVOLVENS 2hl HAPLOPHRAGIIO I DES HALTER I
IS? PLECTOFROHDlCULARlA SP3 2h2 KRRRERIELLA APICULRRIS
184 GVROIDINA OCTOCAMERATA 263 AMMOBACULITES AFF POLVTHALRMUS
187 CIBICIDOIDES GRANULOSA 264 KARRERIELLA CONVERSA
188 PLEUROSTOMELLA S P I 265 ASTERIGERINA GURICHI (PEAK)
I90 ANOMALINOIDES ACUTA 26b GLOBOROTALIR PUNCT ICULATA
!91 'GLOBIGERINA' IFF. H 1 6 6 I N S I 2h7 GLOBOROTALIA HIRSUTA
191 PLANOROTALITES CHAPMAN1 268 GLOBOROTdLlA RFF KUGLERI
196 CSANGULARIA SP4 267 NEOGLOBQUADRINA ATLANTICA
201 SEISMIC EVENT 41 270 C I B l C l D 0 IDES GROSS1
202 SEISMIC EVENT 12 271 GLOBOROTALIR INCREBESCENS
203 SEISMIC EVENT 13 212 GLOBOQUADRINA BRROEROENSIS
204 SEISMIC EVEMT 44 273 BULIMINA GRATA
206, EPOMIDES POLYGONUS 274 GAUORVINA PFF HILTERMANNI
210 LOXOSTOMOIDES APPL INAE 275 PARAROTALIA SP2
 123
TABLE4.8

Partial DAT file for Gradstein-Thomas database. Numbers in brackets below well names a r e for rotary
table height and water depth, respectively (M=meters; F=feet). Depths (first column) are followed by
highest occurrences.

Hibernia P-15 Adolphus D-15 Bjarni H-81 Indian Harbour M-52

( M 11.3; 80.2) (F 98.0: 377.0) (F 40.0; 456.0) (F 98.0; 649.0)
255 17 1140 10 2860 16 1740 1 3
275 18 265 1410 71 3360 67 1740 4 5
310 16 1500 218 3460 20 21 1'740 8
410 20 100 1590 16 136 3560 18 69 1890 9 10
550 26 1680 18 3560 70 71 1950 269
620 201 1980 20 4060 15 2090 2 7
695 15 2700 179 4260 24 2130 6 18
720 71 2900 201 4860 25 2460 15 20
915 72 3060 26 5060 34 2460 16
945 69 3660 15 81 5360 29 265 2550 17
960 202 3660 69 5560 42 74 3600 24 25
975 81 4200 24 33 5560 41 32 4140 26 27
1005 27 4200 202 5560 30 264 4140 28
1035 147 4440 259 25 5560 75 5400 259
1075 24 4562 263 5960 57 5590 261
1125 25 32 4920 82 6060 46 5780 30
1125 57 259 4950 85 261 6590 56 6370 260 32
1125 260 5400 203 6970 33
1185 261 5420 147 260 7660 34 35
1195 29 5550 68 7760 263 36
1200 203 5778 32 7760 39
1315 53 263 5896 90 7860 29 40
1345 40 6018 30 7860 41 42
1375 45 6200 49 29 7960 86
1400 204 6646 144 90 8140 37 38
6646 156 37 8230 44
6646 89 8860 45 46
6975 234 8860 47
7596 160 93 9130 49
7917 36 9560 57 54
8020 161 164 9560 50 52
8258 50 230 9940 55 56
8384 54 10090 59
8520 57 56 10230 60 61
8700 55 10230 62
8726 194 95
124
TABLE4.9

SEQ file for 24 wells of Gradstein-Thomasdatabase for Labrador Shelf and Grand Banks.

BTARNI H-81
16 67 20 -21 18 -69 -70 -71 15 24 25 34 29-261 42 -74 -41 -32 30-264
-75 57 46 56-999
CARTIER D-70
16 18 15 21 -70 67 69 24-172 25 259 34 260-261 118 -85 -29-263 46 -42
-32 35 41 -51 54 56 175 -59-999
F'REYDIS B-87
16 181 -67 -21 -18 20 69 -27 15 -70 25 190 -34-206 -42 -74 260 29-261 -45
33 -81 -41 -75-210 -32 211 -85 -94 57 -88 -86 -30 -46 -35 56 54 213 -55 59
-999
GUDRID H-55
10 -17 265 20 -21 -18 -16 24 15 -25 33 259 40 -34 84 -90 -36 37-260-261
29 35 45 -74 42 57 -88 -30 32 46 -50 56 -59 -54 55-999
INDIAN HARBOUR M-52
1 -3 -4 -5 -8 9 -10 269 2 -7 6 -18 15 -20 -16 17 24 -25 26 -27
-28 259 261 30 260 -32 33 34 -35 263 -36 -39 29 -40 -41 -42 86 37 -38 44
45 -46 -47 49 57 -54 -50 -52 55 -56 59 60 -61 -62-999
KARLSEF'NI H-13
228 67 25 41-118 69 260-261 68 -39 53-206 29 86 -30 -63 -34 46-264 230
-44 -42 96 -36 164 -50 52 45 -54 56 55 -62 61-253 258-999
LEIF M-48
228 -77 -10 181 16 -67 15 20 -21 -18 70 69 85 -24 25-238 42 29 260 -34
57 -74-118-263 30 -41 46 -56 -54-999
LEIF E-38
228 -77-270 17 67 -16 18 -21 20-999
SNORRI J-90
77 228 16 67 15 -21 18 25 57-263 -32 -34 29-260 -53 -41 -30 -36 27 -46
118 264 230 86 -63 42 45 56 59 -54-999
HERJOLF M-92
67 18 -15 -20 -16 78 70 25-259 85-145 -71 -40 45 -35-263-261 -34 29 41
-53 -30 -32-264 86 57 54 46 190 47-154 -56 55 60 59-999
BONAVISTA C-99
76 -77 10 17 -16 21 25 -20 18 79 -15 259 24 -26 81 -33 82 83 40 84
-27 29-261 32-263 85 -86 -87-264 41-34 57 88 -42 -90 89 159 -92 -93 -94
56 -50 -30 47 -96 -36 46-999
DOMINION 0-23
177-109-169 11 -9 17 10-117 -78 112 18 179 -16 -15 -71 122 180 26-123-137
14-136 27 20 21-181 201 24 25 34 264-260 -38 259 142 -81 184 -82 -30-146
69-263 202 32 68 187 49-188-147-190-140 29 -40 191-156 151 250-226 36 -44
194 -90 -57 203 50 -47-158 161 -52 -46 37-159-162 196 45-230 164-999
EGRET K-36
17 26 16 20 -21 -18 -71 -15 24 27 -42 202 69 82-999
OSPREY H-84
17 18 -20 15 -16 26-181 81 82 84-147 -69-148 90 -89 -33-187-234 -34-244
52 -51-162-159-166 -50 -93-999
CUMBERLAND B-55
76 228 -1 17 10 -11 -9-109 -71 265 -16 -20 18 15-119 117 219 26 24 25
-259 132 42 261 41 84 29 32 226 144 49 57 -36 90 52 -54 161 -93 -96-151
-164-157 46 -50-159 55 -56-254-194-999
EGRET N-46
11 -16 -18 14 -27 -71 26 -20 202 15 -24 172-999
ADOLPHUS D-50
10 71 218 16-136 18 20 179 201 26 15 -81 -69 24 -33-202 259 -25 263 82
85-261 203 147-260 68 32 40 30 49 -29 144 -90-156 -37 -89 234 160 -93 36
161-164 50-230 54 57 -56 55 194 -95-999
 125

TABLE 4.9 (continued)

HIBERNIA 0-35
17 201 26 18 -20 16 275 24 -71 72 27 140 202 34 -81 203 259 -29 -25 15
-28 57-260-261 204 40 -32 91-999
nYING FOAM 1-13
9 -10 16 71 17 275-265 18-110 70 26 -15 -81 201 24 -20 -27 25 259 202
263 -32 -34 260-261 264 29 -57-203 54 46 36 41 230-999
BLUE H-28
77 1 4 267 269 110 -10 -64 266 124-125 -6-113 122 26 -71 268 -2 147 -27
29-261 -81-150 82 -15-118-138 146 -84 32 -79-172 -53 -68 164-190 42 86-151
33 -94 -57 37 90 -52-999
HARE BAY H-31
228-270 77 1 10 136 16 70 -15 24 18 -20 -25 260-263 259 29-233 -69-118
-32 -81 68 49 41 227 93 -42 -96 50 57 66 -54 55-161 -56 59 253-255 -46
-999
HIBERNIA K-18
201 16 -18 -20 -71 -72 24 -27 15 -34 81 202 259 147 25 -29-260 30 -57-203
32 263 36 -40 -63 45 -91-155-230204-999
HIBERNIA B-08
17 26 18 -20 16 15 -27 -71 72 81 -25 24 146-259 32 -57-147-260-261-263
36 -40 45 63 47-144-194 -54 -91-230 56 55 -61 52 -59 -96-253-999
HIBERNIA P-15
17 18-265 16 20-100 26 201 15 71 72 69 202 81 27 147 24 25 -32 -57
-259-260 261 29 203 53-263 40 45 204-999

drilled more recently. Rotary table height and water depth are given
separately for each well. For the DEP files to be constructed later for the
purpose of automated stratigraphic correlation, rotary table height will be
subtracted so that all depths were measured from sealevel downward. Feet
will be converted to metres.

Only the relative depths of the samples with respect to one another
are used in ranking and scaling. For example, the Adolphus D-15 well has
32 distinct “event levels” for 50 exits. The majority ( = 19 of 32) of these
levels have a single observed exit; there are 10 levels with 2 , 2 with 3, and
1 with 5 exits, respectively. The total number of samples studied exceeded
the total number of event levels because highest occurrences of
microfossils were coded only. The exits in Table 4.8 have the same
numbers as the Foraminifera in Table 4.7. The complete SEQ file for all 24
wells in the Gradstein-Thomas database is shown in Table 4.9.

4.7 Characteristic features of Gradstein-Thomas database

The original reasons for applying probabilistic stratigraphy (see

Gradstein and Agterberg, 1982) may be summarized as follows. It is well
known that the sequence of first and last occurrences of planktonic
foraminiferal species in open marine Cenozoic sediments in the low-
latitude regions of the world is closely spaced and shows a regular order.
As a result, standard planktonic zonations provide a stratigraphic
resolution of 30 t o 45 zones over a time span of 65 x 106y (Blow, 1969;
Postuma, 1971; Berggren, 1972; Stainforth et al., 1975). Although several
Cenozoic taxa are indigenous to mid-latitudes, the absence of many lower-
latitude forms and the longer stratigraphic ranges of mid-latitude taxa
cause stratigraphic resolution t o decrease away from the lower-latitude
belt. In high latitudes (65"N and S), the virtual absence of planktonic
foraminiferal taxa makes standard zonations inapplicable.
The northwest Atlantic margin, offshore eastern Canada, spans the
mid- t o high-latitudinal realms (north of 42") and although there were
temporal northward incursions of lower-latitudinal taxa in Early o r
Middle Eocene times, there is a drastic overall diminution of the number of
biostratigraphically-useful Cenozoic planktonic species (from about 75 to
30) from the Scotian Shelf to the Grand Banks t o the Labrador Shelf. A
change from a deeper, open marine facies in the Paleogene t o nearshore,
shallower conditions in the Oligocene to Neogene (Gradstein et al., 1975;
Gradstein and Srivastava, 1980) also curtails the number of taxa present
in the younger Cenozoic section.
As a consequence, the construction of a planktonic zonation is mainly
applicable t o the southern Grand Banks and Scotian Shelf where 1 2 zones
have been recognized using species of standard zonations which are not too
rare locally t o be of practical value in correlation. Similarly, on the
northern Grand Banks and Labrador Shelf a 7-fold planktonic subdivision
of the Cenozoic sedimentary strata is possible; the regional application is
limited but the zonal markers and associated planktonic species improve
chronostratigraphic calibration for the benthonic zones.
Independently, the Cenozoic benthonic foraminiferal record also
shows temporal and spatial trends in taxonomic diversity and number of
specimens. Calcareous benthonic species diversity and number of
specimens decreases northward from the Scotian Shelf to the Grand Banks
to the Labrador Shelf whereas the early Cenozoic agglutinated species
diversity and numbers of specimens drastically increases on the Labrador
Shelf. This benthonic provincialism is complicated by incoherent
geographic distribution of some taxa, which in part is due to sampling.
 127

Few of the agglutinated taxa, only a dozen out of more than 50

determined, are of biostratigraphic value (Gradstein and Berggren, 1981),
but among the hundreds of calcareous benthonic forms determined, more
potentially locally-useful or widely-known index species occur. As a
consequence of the ecological sensitivity of these bottom dwellers, and
because of the long stratigraphic ranges, facies changes can be expected t o
modify stratigraphic ranges. This is known as the problem of total versus
local stratigraphic range. A s a result, the benthonic stratigraphic
correlation framework based on exits forms the appearance of a weaving
pattern of numerous small and a few large-scale cross-correlations.
Considerable mismatch in correlation is the result of misidentifiation,
reworking, or large differences between local stratigraphic ranges of a
taxon. In addition, some correlation lines only transverse part of the
combined shelves area.
The previous summary provides insight into some of the constraints
on a regional foraminifera1 zonation. The most important additional one is
sampling method. Only samples of cuttings obtained dominantly over 30-
ft. (10-m.) intervals, are available generally from the wells, inferring that
instead of entry, relative range, peak occurrence, and exit, only the exit of
a taxon is known. Furthermore, downhole contamination in cuttings
hinders recognition of stratigraphically-separate benthonic or planktonic
homeomorphs. Other limiting factors are that species occur frequently in
small numbers and that tests usually are reworked in the younger
Neogene section of the Labrador Shelf.
In summary, the Gradstein-Thomas database of Tables 4.7 - 4.9,
shows the following properties, ranked according to their importance with
respect to stratigraphic resolution:
Samples are predominantly cuttings, which forces use of the highest
parts of stratigraphic ranges or of the highest occurrences (tops,
exits), and restricts the number of stratigraphically useful taxa.
There is limited application of standard planktonic zonations, due to
the mid- to high-latitude setting of the study area and the presence of
locally unfavorable facies.
There are minor and major inconsistencies in relative extinction
levels of benthonic taxa.
128

(4) Many of the samples are small which limits the detection of species
represented by few specimens; this contributes to factor (3) and to the
erratic, incoherent geographic distribution pattern of some taxa.

(5) There is geographic and stratigraphic provincialism in the benthonic

record from the Labrador Shelf t o the Scotian Shelf which makes
representation of details in a general zonation difficult.

Despite the limiting factors, it was possible to erect a zonation based

on a partial database. Gradstein and Williams (1976) used four Labrador
Shelfhorthern Grand Banks wells t o produce an %fold (benthonics)
subdivision of the Cenozoic section. Similar stratigraphic resolution and
improved zone delineation was obtained by Gradstein (unpublished) using
9 wells on the Labrador Shelf and northern Grand Banks. Some of the
zones were tentative and their ages not well defined. These initial
subjective zonations were compared to RASC output (Gradstein and
Agterberg, 1982) suggesting that a slightly improved zonation resulted
from the latter method.

Increase of the Cenozoic database through incorporation of more wells

has clarified the broader correlation pattern and increased the number of
chronostratigraphic calibration points based on planktonic foraminifera1
occurrences. It also increased noise in the stratigraphic signal (factors 3
and 4) due t o more stratigraphic inconsistencies and geographic
incoherence of exits.

The RASC method initially was developed in an attempt to optimize

stratigraphic resolution based on all observations that could be employed
for a zonation. Other benefits of using the computer for ranking and
scaling included the following. Obviously reworked highest occurrences of
taxa never were included in the database. Such reworking is apparent
from anomalous, poor preservation of tests relative to the remainder of the
assemblage and from highly erratic stratigraphic position. However, when
the database is large, it is difficult to evaluate the possibility of anomalous
stratigraphic position for all samples in a systematic manner. The
normality test in RASC (cf. Gradstein, 1984; also see Section 6.6 and
Chapter 8) allows comparison of the positions of the events in each section
with those in the optimum sequence of the biozonation. Events that are
either too high or too low in a given section in comparison with their
neighbors are flagged in the normality test. Such anomalies then can be
 129
scrutinized and excluded from the database if they are due t o reworking,
contamination or misidentification.

4.8 Frequency of occurrence of taxa of Cenozoic Foraminifera

along the northwestern Atlantic margin
In the previous section, it was mentioned that samples obtained
during exploratory drilling are small, limiting the chances t h a t
microfossils will be detected if present within a zone. It is reasonable to
assume that many taxa will not be detected at all in a well. It they are
detected, their highest occurrence is likely t o be recorded a t a
stratigraphically lower level. The first kind of statistical analysis
performed in the RASC program simply consists of counting for how many
different sections (or wells) each taxon has been recorded. Table 4.10
shows such counts for the 150 Foraminifera from the 16 wells in the
northern region introduced at the beginning of Section 4.6 (cf. Fig. 4.5).As
many as 110 events listed in Table 4.10 have zero counts. Most of these
occurred in the southern region only. Some numbers with zero counts
represent “dummy” events (see Section 4.6). In total, 56 events occur in a
single well only. The following tabulation shows how many events occur in
1,2,..., 16 wells of the northern region:

Number of wells: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1516

Numberofevents: 56 26 13 14 11 4 5 2 2 3 4 5 2 1 2 0

This is clearly a skew frequency distribution with relatively few

Foraminifera occurring in relatively many wells. The corresponding
frequency distribution for the southern region is:

Number of wells: 1 2 3 4 5 6
Numberofevents: 56 51 29 21 10 6
TABLE 4.10

RASC computer program preprocessingoutput for number of times that successive events occur in a well;
e.g. event 1 occurs in 2 wells and event 2 in 1 well.

TABULATION OF EVENT OCCURRENCES:

DICTIONARY CODE NUMBER VERSUS FREQUENCY OF OCCURRENCE

I- 2 53- 5 105-0 157-2 209-0

2- I 54- 9 106-0 158-1 210-1
3- 1 55- 6 107-0 159-4 211-1
4- 1 56-12 108-0 160-0 212-0
5- 1 57-12 109-2 161-2 213-1
6- I 58- 1 110-0 162-2 214-0
7- 1 59- 6 111-0 163-0 2 15-0
8- 1 60- 2 112-1 164-3 216-0
9- 3 61- 2 113-0 164-0 217-0
10- 5 62- 3 114-0 166-1 2 18-0
11- 4 63- 3 115-0 167-0 219-1
12- 1 64- 2 116-0 168-0 220-0
13- 1 65- 5 117-2 169-0 22 1-0
14- 3 66- 0 118-4 170-0 222-0
15-14 67- 8 119-1 171-0 223-0
16-15 68- 0 120-0 172-0 224-0
17- 7 69- 10 121-0 173-4 225-0
18-15 70- 7 122-1 174-0 226-1
19- 1 71- 6 123-1 175-1 227-0
20-13 72- 0 124-0 176-4 228-5
21-11 73- 2 125-0 177-1 229-0
22- 6 74- 4 126-0 178-0 230-3
23- 1 75- 3 127-0 179-1 23 1-0
24- 9 76- 2 128-0 180- 1 232-0
25-12 77- 4 129-0 181-4 233-0
26- 7 78- 2 130-0 182-2 234-1
27- 8 79- I 131-1 183-0 235-0
28- 1 80- 1 132- 1 184-1 236-1
29-12 81- 4 133-0 185-0 237-1
30-10 82- 4 134-0 186-0 238-1
31-13 83- 2 135-0 187-1 239-0
32- 4 84- 4 136- 1 188- 1 240-0
33- 4 85- 5 137-1 189-0 24 1-0
34-11 86- 5 138-0 190-3 242-0
35- 5 87- 1 139-0 191-1 243-0
36- 7 88- 3 140-2 192-0 244- I
37- 2 89- 2 14 1-0 193-0 245-0
38- 2 90- 5 142-1 194-2 246-0
39- 2 91- 0 143-0 195-0 247-0
40- 5 92- 1 144- 1 196-1 248-0
41-11 93- 3 145-1 197-0 249-0
4~-11 94- 2 146-1 198-0 250-1
43- 5 95- 0 147-2 199-0 25 1-0
44- 3 96- 3 148-1 200-0 252-0
45- 7 97- 0 149-0 20 1-0 253-1
46-12 98- 0 150-0 202-0 254-1
47- 4 99- 0 151-2 203-0 255-0
48- I 100- 0 152-0 204-0 256-0
49- 3 101- 0 153-0 205-0 257-0
50-10 102- 0 154-0 206-2 258-0
51- 2 103- 0 155-0 207-0 259-0
52- 5 104- 0 156-1 208-0 260-0
 131

It should be kept in mind that a taxon, if it occurred in a well, may have

been observed in several samples. Of these, only the depth of the sample
with the highest occurrence was recorded.
Suppose that the number of wells is represented by the index h. It is
useful t o work with cumulative frequencies expressing how many events
occur in h or more wells. The preceding two tabulations then become:

Northern region:
Number of wells: 1 2 3 4 5 6 7 8 9 1 0 1 1 1 2 1 3 1 4 1 5 1 6

Cumulative frequency: 150 94 68 55 41 30 26 21 19 17 14 10 5 3 2 0

Southern region:
Number of wells: 1 2 3 4 5 6

Cumulative frequency: 157 101 60 31 16 6

The largest cumulative frequency is equal to total number of events in the

region considered.
The cumulative distribution provides a simple guide for selecting a
threshold parameter h, in order t o retain only those events that occur in h,
or more wells. It will be seen later that results of ranking and scaling may
become imprecise if they are based on all events including those that occur
in only one or a few wells. The precision of the results increases when only
those events are used that occur in a t least h, wells. The events occurring
in fewer than h, wells are filtered out. For example, by setting k, = 5 for
the northern region, further analysis was restricted to 41 events. For the
southern region, 60 events with k, = 3 were used. Although statistical
results become more precise when the minimum sample size h , is
increased, an increasingly large number of events then is deleted. The
stratigrapher must make a judicious choice of h, taking care that not too
much information is lost. It is possible that certain key fossils , important
for establishing a regional biozonation, occur in one or a few sections only.
In the RASC method, such special fossils can be coded as “unique” events.
132

These occur in fewer than h, sections. Although unique events are not used
for ranking and scaling, they are inserted later on the basis of their
superpositional relations with other events in the one or more sections
containing them.
The study of the frequency distribution of the events in a region,
selection of the threshold parameter h, and definition of unique events
belong t o the preprocessing module of the RASC computer program.
During this stage, the user should also identify possible “marker
horizons”. These are stratigraphic events with positions that can be coded
with certainty in the h, or more sections containing them. Marker horizons
(e.g. bentonite layers or seismic events) will receive more weight than
other events in the scaling part of RASC.

4.9 Artificial datasets based on random numbers

The Gradstein-Thomas database introduced in the previous sections

is characterized by the fact that it has information on many microfossils
and most of these occur in relatively few sections. Ranking and scaling are
based on superpositional relations between stratigraphic events. If there
are n events in total, the number of pairs of events is n(n-1)/2. For
example, n= 101 results in 5050 pairs. It means that there are fifty times
as many pairs of events as there are individual events. It will be seen in
Chapters 5 and 6 that the frequency distributions for pairs of events in the
Gradstein-Thomas database have smaller frequencies and are even more
skewed than the frequency distributions for counts of events shown in the
previous section.
In order t o test the statistical models for ranking and scaling to be
developed in later chapters it is desirable to have “complete” artificial
datasets in addition to the real datasets. Such artificial datasets can be
obtained from random numbers. In this section, random normal numbers
will be used. In general, it is most convenient to obtain these by means of a
pseudo-random number generator on a computer. Table 4.11 shows how
artificial sequences of three events (A, B and C) can be created from
random normal numbers. The first three columns of Table 4.11 are
random normal numbers from Dixon and Massey (1957). Each number is
a realization of the same random variable X with “normal”, Gaussian
distribution and mean (or expected value) E ( X ) = 2 and variance
Var(X) = 1. By subtracting 1from the numbers in column 1and adding 0.5
 133

TABLE 4.11

Artificial sequences of events A, B and C created from random normal numbers with E(X) = 2 and Var
( X ) = l taken from Table A-23 of Dixon and Massey (1957). Event “Distances” were obtained by
subracting I from random normal numbers in column 1, maintaining column 2, a n d adding 0 . 5 to
random normal numbers in column 3.

Random Event “Distances”

Normal Numbers

1 2 3 A B C Sequence
2.422 0.130 2.232 1.422 0.130 2.732 BAC
0.694 2.556 1.868 -0.306 2.556 2.368 ACB
1.875 2.273 0.655 0.875 2.273 1.155 ACB
1.017 0.757 1.288 0.017 0.757 1.788 ABC
2.453 4.199 1.403 1.453 4.199 1.903 ACB

2.274 1.767 1.564 1.274 1.767 2.064 ABC

3.000 1.618 1.530 2.000 1.618 2.030 BAC
2.510 2.256 1.146 1.510 2.256 1.646 ACB
1.233 2.085 2.251 0.233 2.085 2.751 ABC
3.075 1.730 2.427 2.075 1.730 2.927 BAC

1.344 -0.095 2.166 0.344 -0.095 2.666 BAC

1.246 3.860 1.253 0.246 3.860 1.753 ACB
0.889 2.299 2.458 -0,111 2.299 2.958 ABC
1.154 1.401 1.935 0.154 1.401 2.435 ABC
3.031 1.048 0.719 2.031 1.048 1.219 BCA

0.534 1.155 1.705 -0.466 1.155 2.205 ABC

2.230 3.096 0.045 1.230 3.096 0.545 CAB
2.355 1.761 1.816 1.355 1.761 2.316 ABC
1.461 0.947 0.717 0.461 0.947 1.217 ABC
3.034 1.778 2.122 2.034 1.778 1.622 BAC

2.761 0.473 3.726 1.761 0.473 4.226 BAC

1.961 0.965 1.481 0.961 0.965 1.981 ABC
2.639 4.010 1.915 1.639 4.010 2.415 ACB
1.349 2.225 0.644 0.349 2.225 1.144 ACB
2.959 2.797 4.635 1.959 2.797 5.135 ABC

ACB
CAB
CAB
ABC
ABC
134

TABLE 4.12

Sequences of artificial stratigraphic events A, B and C generated from random normal numbers for
subsamples 1 to 5. Sequences for subsample 1 are same as those shown in last column ofTable 4.11.

I 2 3 4 5
BAC ACR CBA BAC A BC
ACE ACB ACB ACR A BC
ACB RAC ABC ACB A BC
ABC ABC ACB ACB ACB
ACR CAB BAC ACR CAB
ABC CAB CBA ABC ACE
BAC ABC BAC ACE A BC
ACB BCA ACB ARC A BC
ABC ACR ACB ACR ACR
BAC BAC ACE ABC A BC
BAC CBA ACR ARC AC B
ACR ACR ABC ACR BAC
ABC ABC ACB ABC ABC
ABC CBA ACE ARC A BC
BCA ACB ACR BAC ABC‘
ABC BAC ABC BAC CBA
CAB BCA ARC ABC A BC
ABC ABC CAB ABC ACR
ABC ABC ABC BAC A BC
BAC ACB ACB ABC ACR
RAC ACB ABC RAC RAC
ABC ABC ABC ACR CAB
ACB ABC ACE ACB BAC
ACE ABC ACR CRA ARC
ABC CAB ACE ACB A BC
ACE ABC ACR ABC CAB
CAR CAB ARC BAC A BC
CAB BAC ABC BAC ACE
ABC BAC BAC ARC A BC
ABC ACE RCA ACR A BC

t o the numbers in column 3, artificial “distances” along the real line were
created for the events A, B and C which are regarded as realizations of the
normal random variables XA, XB and Xc, respectively.
On the average, the random numbers for events A, B and C occupy the
positions E(XA)= 1.0, E(XB)= 2.0, and E(Xc) = 2.5 which follow one
another along the real line. Consequently, their expected or average
“optimum” sequence is ABC. Each event, however, has variance equal to
one. This implies, that in the realizations, simulating separate
stratigraphic sections, A may be following B or C instead of preceding
them. Thirty “observed” sequences for sections are shown in the last
 135

column of Table 4.11. The artificial sequences are of nine different types
with the following frequencies:

Sequence: ABC ACB BAC BCA CAB CBA

Frequency: 12 8 6 1 3 0

The optimum sequence is observed in 12 of the 30 sections. Because

E(Xb)=2 AND E(Xc)=2.5 are closer together on the real line than
E(XA)= 1 and E(XB)= 2, it is expected that A in the sections precedes B
more frequently than that, for example, B is followed by C. For
frequencies of pairs of events,
Sequence: AB BA AC CA BC CB
Frequency: 23 7 26 4 19 11

It can be attempted, by statistical modelling, t o estimate the optimum

sequence (ABC) and also the relative positions of E(XA),E(XB)and E(Xc)
along the real line from the frequencies of observed sequences in the
sections. Normally such experiments are carried out on a large scale using
a pseudo-random number generator on a computer. An advantage of
computer simulation experiments similar to the experiment of Table 4.11
is, that predictions can be compared to true values, e.g. t o E(XB-XA)= 1.0,
E(XC-XA)= 1.5, E(XC-XB)=0.5. The statistical techniques for making
these predictions will be further developed in later chapters.
The experiment of Table 4.11 was repeated on other random normal
numbers listed in Dixon and Massey (1957, p.452-453) with the resulting
sequences shown in Table 4.12. The final column of Table 4.11 is the first
column of Table 4.12. In this new table, the previous experiment is
regarded as the first subsample for a set of five experiments, all with
E(XA)= 1, E(XB)= 2, E(Xc)= 2.5 and Var(XA)= Var(XB)= Var(XC)= 1. In
the first subsample, the frequencies of the ordered pairs BC and CB were
19 and 11, respectively. The relative frequency of BC, therefore, is
(19/30= )0.633. The set of relative frequencies for all subsamples is
TABLE 4.13

Sequence file with artificially created superpositional relations for 20 events (numbered 1 to 20) in
25 sections. The interval between expected positions of the events along the linear scale was set equal
to 0.5.

1 2 5 4 3 6 10 8 9 11 13 14 12 15 7 17 16 18 19 20

1 4 3 2 7 8 9 6 11 5 12 13 10 15 18 19 16 14 17 20

3 1 2 4 5 6 10 8 7 9 12 11 13 15 16 14 17 18 19 20

5 3 1 2 4 7 6 8 9 10 12 11 13 14 18 19 16 15 17 20

2 1 3 5 6 4 7 8 9 12 10 13 11 14 15 16 19 17 20 18

3 4 5 2 1 6 11 9 7 10 12 8 16 15 14 13 17 18 20 19

2 3 4 1 7 6 9 10 5 12 8 13 14 15 11 16 18 17 19 20

1 3 5 4 9 6 2 7 11 12 8 10 13 16 15 14 17 19 18 20

1 8 3 2 4 6 9 5 12 7 10 11 14 13 15 16 18 17 20 19

2 3 4 1 8 7 6 5 10 12 14 16 11 13 9 15 17 18 19 20

1 5 6 2 3 4 8 7 9 13 10 14 16 11 12 15 17 18 19 20

1 4 6 2 3 5 8 7 9 13 11 14 10 12 15 17 18 16 19 20

2 4 1 5 3 11 6 7 9 8 10 13 14 12 16 15 17 18 19 20

6 3 1 4 2 5 7 8 14 9 11 12 15 16 10 13 17 18 19 20

3 4 2 1 5 7 6 8 9 12 10 11 14 13 16 17 15 19 18 20

3 1 7 6 2 5 4 8 10 15 12 9 13 14 11 17 16 20 19 18

1 2 4 5 7 3 8 6 14 10 9 11 16 12 13 19 18 17 15 20

2 1 4 3 8 6 5 7 9 11 15 14 12 13 10 16 17 20 18 19

1 2 4 7 3 5 6 9 10 11 8 18 13 12 14 15 16 17 19 20

'2 1 4 3 6 5 7 11 10 9 8 14 15 16 12 13 18 17 19 20

3 1 5 4 10 6 2 7 8 11 9 12 14 16 13 17 15 18 19 20

1 2 5 3 4 6 8 7 9 11 10 15 14 13 12 16 19 17 18 20

1 5 4 3 6 2 8 7 11 9 12 10 16 14 17 15 18 13 19 20

2 1 7 3 6 5 4 8 13 12 9 10 11 16 18 20 14 15 19 17

4 1 3 2 8 6 5 7 11 9 13 10 12 16 14 15 17 18 20 19
 137
TABLE 4.14

Sequence file with artificially created superpositional relations for 20 events (numbered 1 to 20) in
25 sections. The interval between expected positions of the events along the linear scale was set equal
to 0.3.

5 1 4 2 10 3 6 8 11 9 15 13 14 17 12 16 7 19 18 20

1 4 3 7 2 8 9 11 6 12 13 18 15 5 10 19 16 20 17 14

3 1 2 4 5 6 10 12 8 9 11 7 16 15 13 17 14 18 19 20

5 3 1 7 6 4 2 9 8 10 12 13 11 14 18 19 17 20 16 15

2 1 3 5 6 8 7 12 9 4 10 14 13 19 15 11 16 17 20 18

3 4 5 11 9 2 6 1 7 10 12 16 15 14 8 17 13 18 20 19

2 3 4 7 1 10 9 6 12 13 15 14 5 8 16 18 11 17 19 20

1 9 3 5 4 6 2 11 7 12 10 16 8 13 15 14 19 17 18 20

8 3 1 2 4 6 9 12 5 10 7 14 11 15 13 16 18 17 20 19

2 3 4 8 7 1 6 10 5 14 12 16 15 13 11 17 9 18 19 20

1 5 6 2 3 8 7 13 4 9 16 14 10 11 12 17 15 18 19 20

1 4 6 5 3 2 8 7 14 13 9 17 11 15 10 12 18 20 19 16

2 4 5 11 3 1 9 6 7 8 13 10 14 16 12 15 17 18 19 20

6 3 4 1 2 5 14 7 8 11 9 16 12 15 17 13 10 18 19 20

3 4 2 1 5 7 12 9 8 6 11 10 14 16 13 17 19 15 18 20

3 1 7 6 5 15 2 10 8 4 14 12 13 9 11 17 16 20 19 18

1 4 7 2 5 14 8 6 3 10 16 11 9 19 12 18 13 17 15 20

2 8 4 1 3 6 7 5 9 11 15 14 12 13 20 18 16 17 19 10

7 1 4 2 5 3 6 9 18 10 11 13 8 12 14 15 16 17 19 20

2 4 1 6 7 3 5 11 14 10 9 8 16 15 18 17 12 13 19 20

3 10 1 5 6 4 7 2 8 11 9 14 12 16 17 13 15 18 19 20

1 2 5 3 4 8 6 7 9 11 15 10 14 13 19 12 16 17 18 20

5 1 4 6 3 11 2 8 7 9 12 16 17 18 14 10 15 13 19 20

2 7 6 1 3 5 13 8 12 4 16 9 10 20 18 11 14 19 15 17

4 3 8 1 2 6 5 11 7 9 13 12 10 16 14 17 15 18 20 19
TABLE 4.15

Sequence file with artificially created superpositional relations for 20 events (numbered 1 to 20) in
25 sections. The interval between expected positions of the events along the linear scale was set equal
toO.l.

5 10 4 2 1 11 17 15 14 8 9 13 6 3 16 12 19 20 18 7

1 4 7 18 11 19 9 13 8 12 3 15 20 2 6 16 17 10 14 5

3 4 1 2 6 LO 12 5 16 11 15 8 9 13 7 17 18 19 20 14

5 7 3 6 9 1 4 8 18 10 19 2 12 13 14 20 11 17 16 15

2 5 12 1 3 19 8 6 7 9 10 15 14 20 16 13 17 4 11 18

11 16 9 5 4 3 10 12 6 15 7 17 2 14 18 1 20 13 19 8

10 15 9 3 7 12 4 2 13 14 6 16 18 1 17 8 5 11 19 20

9 1 5 6 3 4 11 16 12 19 7 15 2 10 13 17 14 18 8 20

8 12 3 9 6 1 15 14 4 2 16 10 13 18 11 7 17 5 20 19

2 8 3 4 7 16 14 10 6 12 15 1 17 13 5 19 18 20 11 9

5 6 1 13 16 8 7 14 9 2 3 4 10 17 12 18 19 11 15 20

4 1 6 5 3 8 2 17 14 13 20 15 19 18 11 7 9 16 12 10

11 4 5 2 9 13 8 7 3 6 10 1 14 16 17 18 19 15 12 20

6 3 14 4 16 11 17 5 15 2 8 1 7 12 9 19 18 20 13 10

3 4 12 5 7 2 9 14 8 1 16 19 17 11 6 10 13 15 18 20

3 1 15 7 6 10 14 8 13 5 12 20 17 2 I 16 11 19 9 18

14 7 16 4 1 19 8 5 2 10 6 13 11 12 17 3 9 13 20 15

2 8 15 20 7 4 6 11 14 9 i9 5 18 3 17 1 13 16 12 10

18 7 4 5 1 9 10 11 2 6 13 3 12 14 16 17 15 8 20 19

2 4 14 1 11 6 7 16 10 15 9 5 3 8 18 17 19 20 13 12

10 3 5 6 7 1 4 8 11 16 14 17 12 9 2 19 18 15 13 20

5 1 2 3 4 8 15 14 11 6 7 19 13 9 10 16 18 17 12 20

5 11 6 4 1 3 8 18 16 17 9 7 12 2 14 15 19 20 10 13

7 13 2 20 6 16 12 18 5 8 3 1 19 10 9 4 11 14 15 17

8 4 11 13 3 6 16 5 17 9 1 2 18 12 7 15 14 10 20 19
 139

S u bsample: 1 2 3 4 5
Relative frequency: 0.633 0.533 0.433 0.600 0.633

The average relative frequency is 0.5667. One might suspect that the
average is a better estimate of the “true” population value because it is
based on a sample that is five times larger. For this example, this
assumption is not correct, because the true relative frequency is
W0.5N2) = 0.638. In the latter expression, CD represents the fractile of the
normal distribution in standard form (see later). In general, if the interval
between the mean positions of two events along the real line is written as
D (D=0.5 for the interval between B and C in the example), then the
population is equal t o Q(DN2).
Tables 4.13 to 4.15 form an artificial database consisting of three SEQ
files for 20 events in 25 sections. The same set of 20x25=500 normal
random numbers was used for each SEQ file. The events are numbered 1
to 20. Because their mean positions follow one another along the real line,
the optimum sequence is also 1to 20 for each SEQ file. The 20 events were
given expected values that are equally spaced. The spacing along the real
line was 0.5,0.3and 0.1 for Tables 4.13,4.14and 4.15, respectively.
Relative frequencies for the order of pairs of consecutive events in
Table 4.13 are similar to those for B and C in Table 4,12, because the
interval D between mean positions is equal to 0.5 in both situations. For
example, the relative frequencies for the first five ordered pairs in Table
4.13 are

Sequence: 12 23 34 45 56
Relative frequency: 0.640 0.520 0.600 0.600 0.560

The average of these five relative frequencies is 0.584. The population

average of 0.638 (see before) would be increasingly closely approximated
by the sample average, if the number of ordered pairs in the sample is
enlarged. One of the advantages of computer simulation experiments is
that the deviations between estimates of parameters based on relatively
small samples and the parameters themselves can be systematically
studied. As pointed out before, the true values of parameters generally are
not available for comparison in real world applications.
This Page Intentionally Left Blank
 141

CHAPTER 5
RANKING OF BIOSTRATIGRAPHIC EVENTS

5.1 Introduction

The purpose of the ranking techniques to be discussed in this chapter

is t o order, for a region, a number of biostratigraphic events for which the
observed superpositional relations in individual stratigraphic sections are
mutually inconsistent. During the 1960s and 1970s, several methods
already were developed to eliminate such inconsistencies in a systematic
manner (Shaw, 1964; Hay, 1972; Rubel, 1978; Davaud and Guex, 1978;
Edwards and Beaver, 1978; for reviews, see Hay and Southam, 1978; and
Brower, 1981). The order obtained for a region after application of a
ranking technique will be called an optimum sequence.

The techniques to be introduced in this chapter and the next (scaling)

show similarity t o the techniques known as “ranking” of objects in
mathematical statistics (cf. David, 1988). According t o Kendall (1975), a
number of individuals are ranked when arranged in order according to
some quality which they all possess to a varying degree. The arrangement
as a whole is termed a ranking in which each member has a rank. An
important difference between the ranking of objects on the basis of their
characteristics and the ranking of stratigraphic events on the basis of
superpositional relations is that, generally, only subsets of all
stratigraphic events are observed within individual sections. These
subsets of stratigraphic events may have sizes which are much smaller
than the total number of events considered for the study region.

In this chapter and the next, ranking and scaling techniques will be
illustrated using the Hay example introduced at the beginning of the
previous chapter. In this example, there are 10 stratigraphic events and 9
sections (see Fig. 4.2; Tables 4.1 and 4.3). The preprocessing of the RASC
computer program begins with a tabulation of the number of stratigraphic
sections in which each event occurs. For the Hay example, this gives:
142

Numberofsections: 8 8 6 7 9 4 7 5 9 6

The following frequency distribution of the stratigraphic events is

obtained from this initial tabulation:

Number o f sections: 1 2 3 4 5 6 7 8 9

Frequency of events: 0 0 0 1 1 2 2 2 2

Curnulativefrequency: 10 10 10 10 9 8 6 4 2
As explained previously (Section 4.81,this frequency distribution is
helpful in selecting the threshold parameter h, which is set to retain only
those events that occur in h, or more wells. For the Hay example, all
events occur in at least 4 sections. Initially, we will set k,= 1 (Default
value for h, in micro-RASC, see Chapter 10) so that all events will be
retained for further analysis.

5.2 Hay’s original method

Hay (1972) began constructing a n optimum sequence from the
stratigraphic information of Figure 4.2 by modifying the subjective
sequence in column 1 on the right side of this diagram. While ignoring
coeval events, Hay counted how often each of the 10 events was observed t o
occur above each of the other events. The resulting counts and
corresponding sample sizes are shown in Figure 5.1A. Dividing a count by
its sample size produces a relative frequency. Because the initial
subjective sequence is not very different from the optimum sequence
(column 2 of Fig. 4.2), most relative frequencies in Figure 5.1A are greater
than 0.5 if they occur above the diagonal consisting of black boxes. Every
relative frequency in the upper triangle of Figure 5.1A has a counterpart
in the lower triangle. Together the relative frequency and its counterpart
add to one and, consequently, most relative frequencies below the diagonal
are less than 0.5.
The optimum sequence is determined by re-evaluating the relations
of all pairs which show a fraction greater than 0.5 in the lower right half of
the matrix. Inspection of the matrix reveals (see Hay, 1972, p. 262) that V
and 8 should be reversed, the number in the appropriate square in the
 143

upper left hand part of the matrix being 1/4, which is less than 0.5. After
making this correction it can be seen that S should come below both 9 and
V, these relationships being expressed by the fractions 0/5 and 1/4,
respectively. Finally, it is evident that the position of in the sequence
needs to be changed because its relation to 6 is 1/4, t o V is 0/5, to q is 1 6 ,
and to < is 1/5. It must come below any of these symbols, and, in fact,
became the lowest event in Hay’s original optimum sequence shown in
column 2 of Figure 4.2. The revised matrix using Hay’s optimum sequence
is shown in Figure 5.1B. All values greater than 0.5 now are in the upper
left part of the matrix. Note that both the upper part and the lower part
contain fractions equal to 0.5. These occur in pairs and signify events that
are coeval “on the average”.
Before or after creation of the optimum sequence, every fraction in the
matrix can be tested for statistical significance by comparing it t o 0.5
using the binomial frequency distribution model as explained in Section
3.2. Figure 5.2 shows the difference between 1 and the cumulative
probability P, ( h , R ) that an event occurs h times above another one in a
sample of pairs of events with size R . If 1-P, ( h , R ) exceeds 0.95, the

Fig. 5.1 (A) Matrix for the relations of biostratigraphic events in Fig. 4.2. The number (N)in the lower
right of each square is the number of sections in which the pair of events is separable. The number ( n ) in
the upper left of each square is the number of times the event on the bottom row occurs below the event
on the left side. The sequence from lowest to highest on the bottom and left side of the matrix is that
shown in column (1) on right side of Fig. 4.2. (B)Revised matrix in which the ratio nlN has been
rearranged so that all values greater than 3 are in the upper left part of the matrix. The lowest-highest
sequence along the bottom and left side of the matrix now represents Hay’s original optimum sequence
also shown as column (2) on right side of Fig. 4.2 (after Hay, 1972).
144
fraction klR is greater than 0.5 with a probability of 95 percent. The
hypothesis of nonrandom average superpositional relationship can only be
accepted for 6 of 45 pairs of events. These are 6 of nine pairs involving the
event W which occurs a t or near the top of all (9) sections (A t o I in
Fig.4.2). In total, two of the values in Figure 5.2 exceed 0.99 They
correspond to the facts that (1)W occurs above in eight sections, and
(2) W occurs above < in eight sections. These two superpositional
relations are statistically significant with a probability of 99 percent.
The binomial model has a drawback for testing whether or not the
observed superpositional relation of two events is random, because it
ignores the relations of these two events with all other events. For
example, the binomial test of Figure 5.2 suggest that W occurs above @.
On the other hand, the fact that A occurs above cD in 4 out of 4 sections
would not be statistically significant, because the sample size is too small.
However, W and A occur near the top in all sections. In those sections
where they coexist, each occurs above the other one 3 out of 6 times. This
would suggest that, although the relation between W and A remains
undecided, both events probably occur above a. The relations between
these three events are shown graphically in Figure 5.3A. If in addition t o

- 5.2 Values of 1-Pwhere P reoresents the orobabihtv that the seauential relation between two events
Fig.
in nonrandom (cf. Eq. 3.2 for cumulative probability of binomial probability with p = 0 . 5 ; after Hay,
1972).
 145

the relations between these three events (W, A and cp), their relations with
a fourth event (V) are also considered, the probability that A occurs above
is further increased (see Fig. 5.3B). A multivariate statistical test which
considers all pairs of events simultaneously and is not subject t o the
drawback of the binomial test of considering pairs of events in isolation,
will be developed in the next chapter on scaling.

5.3 Algorithmic version of Hay’s original method

It is obvious t h a t the method of the previous section can be
programmed for a digital computer. Slightly different versions have been
described in Worsley and Jorgens (19771, Blank (1979), and Agterberg and
Nel (1982a). The following changes help t o make Hay’s method more
general.

1. Choice of initial sequence

Instead of an initial subjective ranking (e.g. column 1 in Fig. 4.2), one
of the sections, if necessary supplemented by information from other
sections, can be used as the starting point. Use of Section A in the Hay
example gave the event numbers 1 to 9 in Tables 4.1 and 4.3. Only the
event A (LO Discolithus distinctus) does not occur in Section A. It was
assigned t h e number 10. While n u m b e r i n g I moved i n t h e
stratigraphically upward direction. However, this decision was arbitrary.

Fig. 5.3 Diagrams to illustrate superpositional relations between (A) three events and (B) four events in
the Hay example. Although A and ID both occur in only 4 sections, their superpositional relation is
probably nonrandom because of their relations with other events.
TABLE 5 . 1

A. F-matrix of frequencies of events occurring above or below one another in the sections. The events for
the Hay example a r e labelled 1 to 10 as in Tables 4.1 and 4.3. B. R-matrix of frequencies ofcoexistence of
two events in the same section. Coeval events also were counted.

A I 2 3 4 5 6 7 8 9 I0

1 x 4 1 1 2 0 2 0 0 0
2 1 x 2 2 1 0 1 0 0 0
3 1 2 x 0 1 0 1 0 0 0
4 4 2 3 x 3 0 3 1 1 1
5 3 3 3 1 x 0 3 0 0 0
6 2 2 2 2 2 X l l O O
7 4 4 3 2 3 1 x 0 0 0
8 5 5 4 3 5 1 4 x 0 0
9 8 8 6 6 9 4 7 5 x 3
1 0 4 4 3 3 4 1 4 2 3 x

0 I 2 3 4 5 6 7 8 9 10

1 ~ 7 5 6 8 3 6 5 8 5
2 7 x 6 6 8 4 6 5 8 5
3 5 6 x 6 6 4 5 4 6 4
4 6 6 6 x 7 5 6 4 7 5
5 a ~ f i 7 ~ 4 7 5 9 6
6 3 4 4 4 4 x 3 2 4 2
7 6 6 5 6 7 3 ~ 5 7 5
8 5 5 4 4 5 2 5 x 5 3
9 8 8 6 7 9 4 7 5 x 6
I0 5 5 4 5 6 2 5 3 6 ~

One could have started by numbering A as 1,followed by W (HI Discoaster

tribrachiatus) as 2 , then moving further downward in Section A.

2. Matrix notation
While arranging the information in matrix form, it is customary to
number the rows from left t o right and the columns from top to bottom.
Table 5.1A shows the so-called F-matrix of frequencies which are similar
t o the counts shown previously in Figure 5.1. The corresponding sample
sizes for frequencies of co-existence of two events in the same section are
shown in Table 5.1B. Note that the main diagonal goes from the top left to
the bottom right side in Table 5.1 .
As already stated in Section 4.3, SEQ files, such as the one shown in
Table 4.3A, normally are for the stratigraphically downward direction
 147

( = direction of drilling exploratory wells in sedimentary basins).

Table5.1A corresponds to Table 4.3A in the following sense. Each
frequency in Table 5.1A indicates how often the event labelling its column
follows the event labelling its row when moving from the left t o the right
through all the rows of Table 4.3. For example, the first element in the
first row of Table 5.1A (after the x on the main diagonal) is equal t o 4. This
means that event 2 (column label) follows event 1four times in Table 4.3A.
The corresponding sections, in which event 2 is stratigraphically below
event 1,are C, D, E and I.

TABLE5.2

A. S-matrix of scores obtained by adding half of the frequencies of ties (shown in Table 5.2B) to the
frequencies of the F-matrix (see Table 5.1A). B. T-matrix of frequencies of ties.

A 1 2 3 4 5 6 1 8 9 10
1 x 5.0 2.5 1.5 3.5 0.5 2.0 0.0 0.0 0.5
2 2.0 x 3.0 3.0 3.0 1.0 1.5 0.0 0.0 0.5
3 2.5 3.0 x 1.5 2.0 1.0 1.5 0.0 0.0 0.5
4 4.5 3.0 4.5 x 4.5 1.0 3.5 1.0 1.0 1.5
5 4.5 5.0 4.0 2.5 x 1.0 3.5 0.0 0.0 1.0
6 2.5 3.0 3.0 3.0 3.0 x 1.5 1.0 0.0 0.5
I 4.0 4.5 3.5 2.5 3.5 1.5 x 0.5 0.0 0.5
8 5.0 5.0 4.0 3.0 5.0 1.0 4.5 x 0.0 0.5
9 8.0 8.0 6.0 6.0 9.0 4.0 7.0 5.0 x 3.0
10 4.5 4.5 3.5 3.5 5.0 1.5 4.5 2.5 3.0 x

B 1 2 3 4 5 6 7 8 9 10
1 x 2.0 3.0 1.0 3.0 1.0 0.0 0.0 0.0 1.0
2 2.0 x 2.0 2.0 4.0 2.0 1.0 0.0 0.0 1.0
3 3.0 2.0 x 3.0 2.0 2.0 1.0 0.0 0.0 1.0
4 1.0 2.0 3.0 x 3.0 2.0 1.0 0.0 0.0 1.0
5 3.0 4.0 2.0 3.0 x 2.0 1.0 0.0 0.0 1.0
6 1.0 2.0 2.0 2.0 2.0 x 1.0 0.0 0.0 1.0
I 0.0 1.0 1.0 1.0 1.0 1.0 x 1.0 0.0 1.0
a 0.0 0.0 0.0 0.0 0.0 0.0 1.0 x 0.0 1.0
9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 x 0.0
10 1.0 1.0 1.0 1.0 2.0 1.0 1.0 1.0 1.0 x
148

3. Incorporation of coeval events

Coeval events were ignored in Figure 5.1 and Table 5.1A. Although
ranking by means of Hay's original method would not be influenced by this
modification, two events which are coeval in a section will be scored by
adding 0.5 t o the two counts for the first event occurring above and below
the second event, respectively. Suppose that the elements of the F-matrix
of Table 5.1A are written as Fij (i = 1, 2, ..., n; j = 1, 2, ..., n ) for n events
( n = 10 in the example). The subscripts i a n d j indicate rows and columns,
respectively. It is noted that these subscripts refer to positions of elements
in a matrix. They do not necessarily coincide with the original code

TABLE5.3

A. P-matrix of relative frequencies obtained by dividing elements of S-matrix by those of R-matrix.

B. Po-matrix of relative frequencies excluding ties.

A 1 2 3 4 5 6 7 8 9 1 0
1 x 5.0/7 2.5/5 1.5/6 3.5/8 0.5/3 2.0/6 0.0/5 0.0/8 0.5/5
2 2.0/7 x 3.0/6 3.0/6 3.0/8 1.0/4 1.5/6 0.0/5 0.0/8 0.5/5
3 2.5/5 3.0/6 x 1.5/6 2.0/6 1.0/4 1.5/5 0.0/4 0.0/6 0.5/4
4 4.5/6 3.0/6 4.5/6 x 4.5/7 1.0/4 3.5/6 1.0/4 1.0/7 1.515
5 4.5/8 5.0/8 4.0/6 2.5/7 x 1.0/4 3.5/7 0.0/5 0.0/9 1.0/6
6 2.5/3 3.0/4 3.0/4 3.0/4 3.0/4 x 1.5/3 1.0/2 0.0/4 0.5/2
7 4.0/6 4.5/6 3.5/5 2.5/6 3.5/7 1.5/3 x 0.5/5 0.0/7 0.515
8 5.0/5 5.0/5 4.0/4 3.0/4 5.0/5 1.0/2 4.5/5 x 0.0/5 0.5/3
9 8.0/8 8.0/8 6.0/6 6.0/7 9.0/9 4.0/4 7.0/7 5.0/5 x 3.0/6
10 4.5/5 4.5/5 3.5/4 3.5/5 5.0/6 1.5/2 4.5/5 2.5/3 3.0/6 x

B 1 2 3 4 5 6 7 8 9 1 0
1 x 4 . 0 ~ 5 1.0/2 1.0/5 2.0/5 0.0/2 2.016 0.0/5 0.0/8 0.014
2 1.0/5 x 2.0/4 2.0/4 1.0/4 0.0/2 1.015 0.0/5 0.0/8 0.0/4
3 1.0/2 2.0/4 x 0.0/3 1.0/4 0.0/2 1.0/4 0.0/4 0.0/6 0.0/3
4 4.0/5 2.0/4 3.0/3 x 3.0/4 0.0/2 3.015 1.0/4 1.0/7 1.0/4
5 3.0/5 3.0/4 3.014 1.0/4 x 0.0/2 3.016 0.015 0.0/9 0.014
6 2.0/2 2.0/2 2.0/2 2.0/2 2.0/2 x 1.0/2 1.012 0.0/4 0.0/1
7 4.0/6 4.0/5 3.0/4 2.0/5 3.016 1.0/2 x 0.0/4 0.0/7 0.0/4
8 5.0/5 5.0/5 4.0/4 3.0/4 5.0/5 1.012 4.0/4 x 0.0/5 0.0/2
9 8.0/8 8.0/8 6.0/6 6.0/7 9.0/9 4.0,'4 7.0/7 5.0/5 x 3.0/6
10 4.0/4 4.0/4 3.0/3 3.0/4 4.014 1.0/1 4.0/4 2.0/2 3.0/6 x
 149

numbers of the events. The resulting modified matrix t o be used here is

the S-matrix shown in Table 5.2A. Also shown are the symmetrical
T-matrix (Table 5.2B) for frequencies T, = Tji of coeval events (or “ties”).
The R-matrix for sample sizes R, = Rji of pairs of events including ties
was already shown in Table 5.1A. Consequently, the scores S,j.tabulated
in the S-matrix, satisfy the equation: Sij=Fij++T,j..

Relative frequencies P,j. with P,j.= S,j./Rij can be formed by dividing

every score by the corresponding sample size in the R-matrix. The
resulting P-matrix for relative frequencies is shown in Table 5.3A.
Suppose t h a t sample sizes without counting ties are denoted a s
Rou =Rij-Tij. For comparison, the relative frequencies POG =F,/Ro, are
shown in the Po-matrix of Table 5.3B. These relative frequencies were
previously shown in Figure 5.1. Note that any attempt t o move all relative
frequencies greater than 0.5 to positions above the main diagonal would
yield identical results which are independent of whether the P-matrix or
the Po-matrix is used. Later (see Chapter 6), it will be shown that there
are advantages t o using P instead of Po when all superpositional relations
between events are considered simultaneously.

4. Order of checking superpositional relations

In Hay’s original example, the order in which events were selected for
comparison with other events was subjective. For a n algorithm, it is
preferable t o proceed in the same way in all applications if possible. The
obvious choice is to begin at the beginning of the first row. For example,
the first comparison then to be made in the S-matrix of Table 5.2A is for
the element S12 = 5 versus S21= 2. Since S12 is greater than S21 it is not
necessary t o reverse the order of events. The next pair of events to be
tested is s13=2.5 and S31=2.5. Again it is not necessary t o reverse the
order, this time because the two matrix elements are equal t o one another.
The next pair is S 1 4 = 1.5, S41=4.5. Because S41>S14 the positions of the
first and fourth rows and columns should be interchanged. Table 5.4A
shows the revised matrix after the interchange. It now is necessary to
return t o the first element of the first row for comparison with its
counterpart, because the new first row is what originally was the fourth
row (with the first element of the original fourth row in the fourth column
of the new first row). The original code numbers are shown in parentheses
in Table 5.4A.
150

TABLE 5.4

Illustration of algorithm for systematic checking of superpositional relations i n Hay method for
constructing optimum sequence. A. Positions of events 1 and 4 were interchanged because in Table 5.2A
the element ( = 1.5)in the fourth column of the first row is less than its counterpart (=4.5)in the lower
triangle of the matrix. Original event code numbers a r e shown in parentheses. B. Positions of events 6
and 4 were interchanged during second iteration. C . Positions of events 9 and 6 were interchanged
during third iteration. D. Final order relation matrix after 22 iterations. This matrix has the property
that all its elements in the upper triangle a r e greater than or equal to their counterparts in the lower
triangle. Elements in the upper triangle equal to their counterparts are underlined in Table 5.4D. The
events corresponding to these elements are coeval on the average. Note t h a t the final (optimum)
sequence is nearly the reverse of the original sequence in Table 5.2because code numbers were assigned
to the events while moving in the stratigraphically upward direction (cf. Tables 4.1 and 4.3).

A 1 2 3 4 5 6 7 8 9 1 0
I41 12) (31 ill 151 161 171 181 19) 1101
-
1141 x 30 45 45 15 I0 35 10 10 15
2121 30 x 30 20 30 10 15 00 00 05
3131 I5 30 25 20 10 15 00 00 05
4111 15 50 25 x 35 05 20 00 00 05
5151 25 50 40 25 .i 10 35 00 00 10
6161 30 30 30 30 30 15 10 00 05
I(7) 25 45 35 25 35 15 x 05 00 05
8(81 30 50 40 30 50 10 45 x 00 05
9191 60 80 60 60 90 40 I0 50 x 30
101101 35 15 35 35 50 IS 45 25 30 x

8 1 2 3 4 5 6 7 8 9 1 0
161 121 131 Ill 151 (41 171 I81 191 (101

1161 x 30 30 25 30 30 15 10 00 05
2(21 10 x 30 20 30 30 I5 00 00 05
301 10 30 x 25 20 I5 15 00 00 05
4(1) 05 50 25 v 35 15 20 00 00 05
5(51 I0 50 40 45 x 25 35 00 00 10
6(41 10 30 45 45 45 x 35 10 10 15
I(7l 15 45 35 40 35 25 x 05 00 05
8(8) I0 50 40 50 50 30 45 x 00 05
91% 40 80 60 80 90 60 70 50 x 30
101101 I5 45 35 45 50 35 45 25 30 x

C I 2 3 4 5 6 7 8 9 1 0
191 121 131 Ill (51 (41 I71 18) (61 110)
1191 x 80 60 80 90 60 I0 50 40 30
2121 00 x 30 20 30 30 15 00 10 05
3(31 00 30 x 26 20 15 15 00 10 05
411) 00 50 25 x 35 I5 20 00 05 05
515) 00 50 40 45 x 25 35 00 10 10
6141 10 30 45 45 45 x 35 10 10 15
7171 00 45 35 40 35 25 x 05 15 05
8181 00 50 40 50 50 30 45 II 10 05
9(61 00 30 30 25 30 30 15 10 x 05
1011Ol 30 45 35 45 50 35 45 25 15 x

D 1 2 3 4 5 6 7 8 9 1 0
I91 1101 (61 181 14 171 151 11) (91 121
119) x 30 40 SO 60 70 90 80 60 80
21101 6 I5 26 35 45 50 45 36 41
3161 00 06 x Q 30 Is 30 25 30 30
4181 00 05 Q x 30 48 50 50 40 50
5(4) I0 I5 I0 10 x 35 45 45 45 30
8(7L 00 05 05 25 x 35 40 35 45
I151 00 LO 10 00 '25 x 45 40 30
8111 00 05 05 00 18 20 35 x Q 5G
9131 00 05 I0 00 15 15 20 2.6 x 9
lIll2) 00 05 10 00 30 15 30 20 30 (i
 151

TABLE 5.5

Optimum sequence output of the RASC computer program. Order of events is same as in Table 5.4D.

Sequence Uncertainty Event Event

Number Range Code Name
1 0-3 9 HI Discoaster tribrachiatus
2 0-3 10 LO Discolithus distinctus
3 2-5 6 LO Rhabdosphaera scabrosa
4 2-5 8 LO Discoaster cruciformis
5 4-6 4 LO Coccolithus solitus
6 5-8 7 LO Discoaster minimus
7 5-8 5 LO Coccolithus gammation
8 7-10 1 LO Discoaster distinctus
9 7-1 1 3 LO Discoaster germanicus
10 8-11 2 LO Coccolithus cribellum

The step of making one interchange because an element in the upper

triangle is less than its counterpart in the lower triangle will be called a n
iteration. Successive checking of the elements in the first row of
Table 5.4A shows that a second iteration is required at the sixth column
because s61>s16. It means that the first and sixth rows and columns
should be interchanged. The result of this second iteration is shown in
Table 5.4B. A s shown in Table 5.4C one can proceed to the ninth column
before the third iteration is required. In Table 5.4C, the situation is finally
reached that none of the elements in the first row is less t h a n its
counterpart in the first column. It means that one can proceed t o the
second row. The first element to be tested now is in the third column. The
fourth iteration consists of interchanging the positions of the second and
fourth rows and columns. In general, once all elements of a given row in
the upper triangle have passed the test of comparing them to their
counterparts in the corresponding column, then it will not be required t o
test them again, although they may be moved to other positions within the
same row during subsequent iterations. Continuation of the algorithm
finally led to the matrix of Table 5.4D, after 22 iterations in total. This is
the so-called final order relation matrix. The order of the events in this
matrix is considered to be the optimum sequence.
152

5. Consideration of events which are coeval o n the average

A number of elements are underlined in Table 5.4D.They belong to

pairs of events which are coeval on the average. In total, there are 6 pairs
of this type. The elements of 5 of these 6 pairs are adjoining the main
diagonal. If the positions of events which are neighbors in the optimum
sequence are interchanged, the sequence remains an optimum sequence
because none of its lower triangle elements exceeds 0.5. For example, if
events 9 and 10, which are in positions 1 and 2 respectively, are
interchanged, all frequencies in the upper triangle remain greater than
their counterparts in the lower triangle. This rule does not apply to pairs
in the optimum sequence which are coeval on the average but are
separated by one or more events with which they are not coeval on the
average. For example, events 6 and 7, which are in positions 3 and 6, are
separated by events 8 and 4. If events 6 and 7 are interchanged, the
resulting sequence is not an optimum sequence because event 7 follows
event 4 in most sections, while event 4 follows event 6 in most sections
containing both events. Consequently, event 7 must follow event 6 in any
optimum sequence.

5.4 Uncertainty ranges for events in the optimum sequence

It is useful t o define an uncertainty range for the events in the

optimum sequence. Table 5.5 shows the RASC output for the optimum
sequence of Table 5.4D. The first column contains the sequence numbers
of the events in the optimum sequence. Column 3 gives the original code
numbers and the names of the events are shown in the last column. The
uncertainty range in the second column of Table5.5 applies to the
sequence number. Its two numbers are less than and greater than the
sequence number, respectively. This range was determined by counting,
for each event, the number of adjoining events with which it is coeval on
the average. For example, because the positions of events 9 and 10 can be
interchanged, and there are no other, similar pairs in the vicinity, their
uncertainty ranges are 0-3. This indicates that the sequence number of
either event could be 1 or 2. It is not possible to decide whether event 9
should come before or after 10 in the optimum sequence. On the other
hand, the uncertainty range of event 4 extends from sequence position 4 t o
6 indicating that its sequence position ( = 5) is not, on the average coeval
with any other event. Although events 6 and 7 are coeval on the average,
 153

it could be established (see before) that event 6 must precede 7 in the

optimum sequence. This type of uncertainty does not show up in the
uncertainty range.
In general, the uncertainty range provides a quick method for
evaluating how firmly an event is positioned between its neighbors in the
optimum sequence. Occasionally, the uncertainty ranges of successive
events interact with one another and the possible positions of the events
are not immediately obvious. For example, in Table 5.5 events 1,3 and 2
have uncertainty ranges 7-10, 7-11 and 8-11, respectively. This means
that event 1 or 3 (but not 2) can occupy position number 8. It also means
that 2 or 3 (but not 1) can have position 10. Although all three events can
occupy position number 9, the preceding conditions imply that 3 must
precede 2. This type of conclusion can be drawn more rapidly by inspection
of the frequencies in the final order relation matrix shown in Table 5.4D.

Three events A, B and C as a group are mutually inconsistent if, on

average, A occurs before B, B before C, and C before A. It will be shown
later that if the superpositional relations of 3 or more events are mutually
inconsistent, it is not possible t o construct an optimum sequence by Hay’s
original method. Neither can then an optimum sequence be obtained by
the algorithm of Section 5.3. A solution can, however, be obtained by
ignoring one or more pairs of scores (Sij and Sji) for events participating in
inconsistencies involving groups of more than two events. In RASC,
ignored pairs of this type will be treated as pairs with equal scores when
the uncertainty range is determined.
In general, the scores Sij and Sji are subject t o a statistical
uncertainty which, in a relative sense, decreases with increasing sample
+
size. Rij ( = S,j. Sj$. If the statistical population from which a sample with
size R,j. is drawn has fixed probability nij that event i is followed by event
j , then the difference between the observed proportion Pij ( = S,j./Ru)and
n,j.is relatively large when Rij is small. Binomial theory can be used to
quantify the frequency distribution of P,j. of which the mean value is nu.
This dependence on sample size implies that the erroneous observation
Sji>Sij (if on the average S,j.>Sji) will be made more frequently when R,j.
is small. In RASC, the user has the option of ignoring pairs of scores if
sample size is less than a selected threshold value m,l. In the previous
example, m,l= 1 so that all pairs were used. However, if one were t o set
m,.=3, two pairs of events with sample size R,j.=2, would be ignored in
Table 5.4D. These are the pairs (10,6) and (6,8), respectively. For
154

determination of the uncertainty range, pairs of events that are ignored

because of the introduction of a threshold value will be treated in the same
way as pairs of events that are coeval on the average. By this method, it is
possible to consider, to some extent, the statistical uncertainty of event
positions in the optimum sequence. Better methods t o express the
statistical uncertainty of the average position of events can be derived
after scaling the events (Chapter 6).

5.5 Other ranking algorithms

In total, 22 iterations were required t o produce an optimum sequence
(Table 5 . 5 ) from the original S-matrix (Table 5.3A). In this section, faster
algorithms will be discussed which lead t o exactly or approximately the
same final product. From a practical point of view, it is not important
which one of the algorithms would be selected for this particular example,
because there is no significant difference in the computing time required.
In other applications, however, hundreds of thousands or more iterations
might be required. Then it may become necessary to switch to algorithms
by means of which an optimum sequence is produced faster. One method
by which the total number of iterations generally can be ranked very
quickly, is to set a tolerance value (b,) greater than zero for the differences
Sji-Sij. In the previous algorithm, an iteration is carried out if S j i - s >
~ 0.
The user can require that an iteration is only carried and if Sji-Si~> b, with
b,>O. The option of making the tolerance 6 , greater than in its default
value, which is equal to zero, is available in the RASC computer program.
This option reduces the computing time required to obtain an optimum
sequence but this accomplished by leaving a variable amount of “noise” in
the result.

Use of transposed order relation matrix

It is obvious that a relatively large number ( = 22) of iterations was

required for the example of Table 5.4 because, initially, the majority of the
scores in the upper triangle were less than their counterparts in the lower
triangle. The transpose of the original S-matrix (Table 5.3A) is obtained
by replacing Sij by SJi (and Sji by S Q ) . The transpose is shown in
Table5.6A. If the algorithm is applied, the first iteration consists of
interchanging events 10 and 9 which occupy the first and second position
 155

TABLE 5.6

A. Transposed S-matrix (cf. Table 5.2A). B. Final order relation matrix obtained after 5 iterations

A I 2 3 4 5 6 7 8 9 10
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

1(1) x 2.0 2.5 4.5 4.5 2.5 4.0 5.0 8.0 4.5
2(2) 5.0 x 3.0 3.0 5.0 3.0 4.5 5.0 8.0 4.5
3(3) 2.5 3.0 x 4.5 4.0 3.0 3.5 4.0 6.0 3.5
4(4) 1.5 3.0 1.5 x 2.5 3.0 2.5 3.0 6.0 3.5
5(5) 3.5 3.0 2.0 4.5 x 3.0 3.5 5.0 9.0 5.0
6(6) 0.5 1.0 1.0 1.0 1.0 x 1.5 1.0 4.0 1.5
7(7) 2.0 1.5 1.5 3.5 3.5 1.5 x 4.5 7.0 4.5
8(8) 0.0 0.0 0.0 1.0 0.0 1.0 0.5 x 5.0 2.5
g(9) 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 x 3.0
lO(10) 0.5 0.5 0.5 1.5 1.0 0.5 0.5 0.5 3.0 x

B 1 2 3 4 5 6 7 8 9 10
(2) (1) (3) (5) (7) (4) (6) (8) (9) (10)
x 5.0 3.0 5.0 4.5 3.0 3.0 5.0 8.0 4.5
2.0 x 2.5 4.5 4.0 4.5 2.5 5.0 8.0 4.5
3.0 2.5 x 4.0 3.5 4.5 3.0 4.0 6.0 3.5
3.0 3.5 2.0 x 3.5 4.5 3.0 5.0 9.0 5.0
1.5 2.0 1.5 3.5 x 3.5 1.5 4.5 7.0 4.5
3.0 1.5 1.5 2.5 2.5 x 3.0 3.0 6.0 3.5
1.0 0.5 1.0 1.0 1.5 1.0 x 1.0 4.0 1.5
0.0 0.0 0.0 0.0 0.5 1.0 1.0 x 5.0 2.5
0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 x 3.0
0.5 0-.5 0.5 1.0 0.5 1.5 0.5 0.5 3.0 x

in the sequence of columns and rows in Table 5.6A. Table 5.6B shows the
final order relation matrix which now was obtained after 5 iterations only.
Table 5.7A is RASC output for the optimum sequence of Table 5.6B.
The original SEQ file for this RASC run was shown in Table 4.3B.
Because proceeding from left to right in this SEQ file corresponds t o
moving in the stratigraphically upward direction, the optimum sequence
of Table 5.7A is upside down. Table 5.7B is identical to Table 5.7A except
for a reversal of the sequence numbers. It is interesting to compare
Table5.7B with the previous result (Table 5.5). The sequence order is
different in 4 places. In 3 of these, the order of a pair of two events was
156

reversed. This possibility is expressed by the uncertainty ranges of the

events which are identical except for event number 10 which has
uncertainty range 8-11 in Table 5.5 and 9-11 in Table 5.7B. This is
because the uncertainty ranges of events 8, 9 and 10 interact with one
another as explained in Section 5.5. The uncertainty range of 8-11 for
event 10 in Table 5.5 is more meaningful than 9-11 in Table 5.7B because
event 10 could occur in position 9 provided it would be followed by event 8
in position 10. This illustrates that for a full appreciation of the
interaction of uncertainty ranges it may be necessary t o inspect the
elements of the final order relation matrix.

Use of a transposed order relation matrix is equivalent to reversing

the direction for coding t h e superpositional relations between
stratigraphic events. Provided that the uncertainty range is considered,
the final optimum sequence is nearly independent of this type of reversal.

Probabilistic ranking

The simple algorithm here termed “probabilistic ranking” was

originally added to the RASC computer program as a “presorting option”
(Agterberg and Nel, 1982a). It resembles a method earlier proposed by
Rube1 (1978) which will be discussed in Section 5.6. It will be shown here
that, for the Hay example, probabilistic ranking produces the same
optimum sequence (Table 5.5) as the algorithm discussed earlier in this
chapter. The problem of cycling due to inconsistencies involving more
than two events (see Section 5.4) is avoided in probabilistic ranking.
Harper (1984) has shown that, in his computer simulation experiments
(see Section 7.41, “presorting” consistently gave better results than the
modified Hay method which is essentially the same as the algorithm of
Section 5.2 with modifications to account for cycling. In Agterberg and
Nel (1982a), it was recommended t o use presorting followed by the
modified Hay method. The new term “probabilistic ranking” reflects that
the algorithm previously termed presorting often produces better results
than the modified Hay method.

Probabilistic ranking consists of replacing the elements S,j. in the S-

matrix by Sij = 1if Sg >Sji, by Sij = O if Sg >Sji and by Sg = 0.5 if Sij = Sji.
Table 5.8 shows the A-matrix with elements A,j. corresponding t o the S-
matrix of Table 5.2A. By ordering the row totals Ai according t o
decreasing magnitude, the optimum sequence of Table 5.9 was obtained.
 157

TABLE5.7

A. Optimum sequence output of RASC computer program corresponding to Table 5 . 6 8 . This result was
obtained by using Table 4.3B as SEQ tile instead of Table 4.3A. B. Reversed optimum sequence of
Table 5.7A. The sequence numbers 1 to 10 for the optimum sequence of Table 5.7A were replaced by new
sequence numbers 10 to 1 .

A. Sequence Uncertainty Event Event

Number Range Code Name
1 0-2 2 LO Coccolithus cribellum
2 1-4 1 LO Discoaster distinctus
3 0-4 3 LO Discoaster germanicus
4 3-6 5 LO Coccolithus gammation
5 3-6 7 LO D i s c o a s h minimus
6 5-7 4 LO Coccolithus solitus
7 6-9 6 LO Rhabdosphaera scabrosa
8 6-9 8 LO Discoaster cruciformis
9 8-11 9 HI Discoaster tribrachiatus
10 8-11 10 LO Discolithus distinctus

B. Sequence Uncertainty Event Event

Number Range Code Name
1 0-3 10 LO Discolithus distinctus
2 0-3 9 HI Discoaster tribrachiatus
3 2-5 8 LO Discoaster cruciformis
4 2-5 6 LO Rhabdosphaera scabrosa
5 4-6 4 LO Coccolithus solitus
6 5-8 I LO Discoaster minimus
7 5-8 5 LO Coccolithus gammation
8 7-11 3 LO Discoaster germanicus
9 7-10 1 LO Discoaster distinctus
10 9-11 2 LO Coccolithus cribellum

The algorithm for sorting events according t o their magnitude is

illustrated in Table 5.10. It consists of the following steps. The event with
sequence number 1 successively was compared with all following events
and its position was interchanged with that of a successor if its magnitude
was less. This automatically brings the event (9) with the greatest row
total (8.5)to the first position in the optimum sequence. The order of 9 and
10 is not changed because they have the same magnitude. When the event
with the largest magnitude is in first position, the algorithm proceeds t o
158

TABLE5.8

A-matrix to denote average superpositional and coeval relations. Method of probabilistic ranking (or
“presortingoption”) applied to Hay example using S-matrix of Table 5.2A as starting point. F-matrix of
Table 5.1A gives same A-matrix. Events will be reordered on the basis of their row totals (At).

1
2
1

x
0.0
2

1.0
x
3

0.5
0.5
4

0.0
0.5
5

0.0
00
6

0.0
0.0
7

0.0
0.0
8

0.0
0.0
9

0.0
0.0
lo
0.0
0.0
I A‘
1.5
10
3 0.5 0.5 x 0.0 0.0 0.0 00 0.0 0.0 0.0 1.0
4 10 05 1.0 x 1.0 0.0 1.0 0.0 0.0 0.0 4.5
5 1.0 1.0 1.0 0.0 x 0.0 0.5 0.0 00 0.0 3.5
6 1.0 1.0 1.0 1.0 1.0 x 0.5 0.5 0.0 0.0 6.0
7 1.0 1.0 1.0 0.0 0.5 0.5 x 0.0 0.0 0.0 4.0
8 10 10 1.0 1.0 1.0 0.5 1.0 x 0.0 00 6.5
9 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 x 0.5 8.5
10 1.0 1.0 1.0 1.0 1.0 1.0 1.0 10 0.5 x 8.5
A, 75 80 80 45 55 30 50 25 05 05 1

carry out similar tests for the second position. In Table 5.10 it is shown
that it took four iterations t o bring event 9 to position 1, followed by five
iterations t o bring event 10 to position 2. Continuation of the algorithm to
find the events for the third and subsequent positions gave the optimum
sequence of Table 5.9 after 31 iterations. The new result is identical t o
that obtained before (Table 5.5). The uncertainty range of an optimum
sequence obtained by probabilistic ranking can be determined by using the
same method as before (see Section 5.4).
As a further experiment, probabilistic ranking was applied using the
SEQ file of Table 4.3B instead the one of Table 4.3A. This is more or less
equivalent t o ranking the events in ascending order using the column
totals Aj of Table 5.8. When the events were first ranked according to
descending order of magnitude of their column totals, reversal of the
resulting optimum sequence gave an optimum sequence identical to the
one shown in Table 5.7 except that event 10 was situated above event 9.
The uncertainty ranges resulting from this experiment were identical t o
those given in Table 5.9.
 159

TABLE 5.9

Optimum sequence output of RASC computer program corresponding to Table 5.8. Events were
reordered on the basis of their row totals.

Sequence Code Row Uncertainty

Number Number Total Range
1 9 8.5 0-3
2 10 8.5 0-3
3 8 6.5 2-5
4 6 6.0 2-5
5 4 4.5 4-6
6 7 4.0 5-8
7 5 3.5 5-8
8 1 15 7-10
9 3 I .o 7 - 11
10 2 1.0 8-11

Missing data in probabilistic ranking

In practice, the S-matrix may contain pairs of zero elements with

S,j.= Sji = 0 because of missing data. The corresponding elements in the A-
matrix then can also be set equal to zero (Ai,.=Aji=O). A distinction
should be made between a zero whose counterpart is equal t o one, and t o a
zero whose counterpart is zero because it belongs to a pair of zeros for
missing information. Suppose that there are Bi zeros of the second type in
the i-th row. The row total E j Aij may be biased ( = t o o small) because one
or more of the missing elements with values equal t o 0.0 in reality could be
0.5 or 1.0. The count Bi can be combined with the possibly biased row total
t o produce the ranking number

A i = (n-1)(EjA ij)(n-l-Bi)-' (5.1)

This is equivalent to rescaling totals for rows with missing information in
such a way that the sum of each Ai and its corresponding column total
remains equal to (n-1).
Table 5.11A (from Agterberg and Nel, 1982a, p. 74) provides an
example of this type of rescaling. Twenty-six highest occurrences of
Cenozoic Foraminifera, each occurring in at least h, = 7 offshore wells
along the northwestern Atlantic margin were subjected to probabilistic
160

TABLE 5.10

Illustration of computer algorithm used in probabilistic ranking to reorder events on the basis of their
row totals in Table 5.8. Final result obtained after 31 iterations is identical to results previously
obtained by Hay method (cf. Tables 5.4 and 5.5).

Iteration I 2 3 4 5 6 7 8 9 10

1 4 2 3 I 5 6 7 8 9 IIJ
2 6 2 3 I 5 4 7 8 9 10
3 8 2 3 I 5 4 7 fi 9 10
4 9 2 3 1 5 4 7 6 8 10
5 1 3 2 5 4 7 6 8 10
6 5 3 2 I 4 7 6 8 10
7 4 3 2 1 5 7 6 8 10
8 6 3 2 I S 7 4 8 10
9 8 3 2 1 S 7 4 6 10
10 10 3 2 I 5 7 4 6 8
11 1 2 3 5 7 4 6 8
12 5 2 3 1 7 4 6 8
13 7 2 3 1 5 4 6 8
I4 4 2 3 1 5 7 6 8
15 6 2 3 1 5 7 4 8
16 8 2 3 1 5 7 4 6
17 1 3 2 5 7 4 6
in 5 3 2 1 7 4 6
19 7 3 2 1 5 4 6
20 4 3 2 1 5 7 6
21 6 3 2 1 5 7 4
22 1 2 3 5 7 4
23 5 2 3 1 7 4
24 7 2 3 1 5 4
25 4 2 3 1 5 7
26 1 3 2 5 7
27 5 3 2 1 7
28 7 3 2 1 5
29 1235
30 5 2 3 1
31 1 3 2

ranking. The ranking numbers of events 26 and 67 are revised row totals.
For this reason, they are not multiples of 0.5 like the other ranking
numbers in Table 5.11A. Reordering the 26 events on the basis of the
ranking numbers gives the optimum sequence of Table 5.11B.
Probabilistic ranking can be regarded as a primitive kind of scaling
method because the events are assigned values along an interval scale.
 161
TABLE 5.11

A . Ranking n u m b e r s A , obtained by method of probabilistic r a n k i n g applied t o 26 Cenozoic

foraminifera1 events which occur ink,= 7 or more wells. Original event numbers a r e shown in column 1.
New ranks obtained from ranking numbers A, a r e shown in the fourth column. B. The ranks a r e shown
in ascendingorder so t h a t events a r e in optimum sequence.

A: Event i A, Rank B Rank Event

15 1 19.5 7 1 17
16 2 24.0 2 2 16
17 3 25.0 1 3 67
18 4 21.5 4 4 18
20 5 20.0 6 5 21
21 6 20.5 5 6 20
24 7 15.5 10 7 15
25 8 15.0 11 8 26
26 9 18.2 8 9 70
27 10 14.0 13 10 24
29 11 11.5 15 11 25
30 12 7.0 19 12 69
31 13 12.0 14 13 27
34 14 10.0 16 14 31
36 15 5.5 20 15 29
41 16 9.0 17 16 34
42 17 8.0 18 17 41
45 18 4.5 22 18 42
46 19 3.0 23 19 30
50 20 2.5 24 20 36
54 21 1.0 25 21 57
56 22 0.0 26 22 45
57 23 4.5 21 23 46
67 24 23.9 3 24 50
69 25 14.0 12 25 54
70 26 17.0 9 26 56

Scaling by the averaging ofprobabilities

Probabilistic ranking gives approximately the same results when the
A-matrix is constructed from the F-matrix instead of the S-matrix. The
162

TABLE 5.12

Ranking numbers obtained by averaging probabilities for the Hay example. See text for further
explanation.

(1) (2) (3) (4) (5) (6)

I 15 5 53 10 42 0 292 0 238
2 14 0 55 7 43 0255 0 163
3 12 0 46 5 32 0261 0 156
4 24 5 51 18 38 0480 0474
5 21 5 60 13 43 0358 0302
6 17 5 30 12 19 0583 0632
7 20 5 50 17 43 0410 0395
8 28 0 38 28 36 0737 0 778
9 56 0 60 56 60 0933 0933
10 32 5 41 28 32 0793 0 875

Sum 242.0 484 194 388

only possible difference between outcomes resulting from these two

procedures would be due to pairs of locally coeval events which are not
considered i n the F-matrix. A difference of this type does not arise when
probabilistic ranking is applied t o the F-matrix of Table 5.1A o r the
corresponding S-matrix (Table 5.2A).

Suppose t h a t for each row in Table 5 . 1 A o r 5 . 2 A , t h e relative

probabilities (shown in Tables 5.3B and 5.3A, respectively) would be added
without first replacing these matrices by the A-matrix. Division of its sum
by (a-1) would give a n average probability for each event. It can be argued
that the probabilities are of variable precision. Their variance is inversely
proportional to sample size ( = number of pairs). This suggests that i t
would be advantageous to compute a weighted average of the probabilities
in each row using the sample sizes a s weights. Multiplication of a
) its sample size R,j. yields the original frequency
probability (e.g. P ~ Jby
(e.g. Sg =P,j.X Rij). Consequently, the suggested best procedure simply
consists of summing the scores in each row of the S-matrix and t h e n
dividing the resulting row sums by the corresponding sums for rows of the
R-matrix.

Table 5.12 shows r a n k i n g numbers obtained by averaging t h e

probabilities P,j. (column 5 ) and Pog (column 6) for the events of the Hay
example, respectively. The average probabilities of column 5 were
obtained by dividing the numbers in column 1 by those in column 2 which
 163

are row totals for the S-matrix (Table 5.2A) and the R-matrix (Table 5.1B),
respectively. The sum of the row totals in column 2 is twice as large as the
sum of the row totals in column 1. The numbers in column 3 of Table 5.12
are row totals for the F-matrix (Table 5.lA). These were divided by the
numbers of column 4 that represent sample sizes for pairs of events after
exclusion of ties (Table 5.2B). The sum for column 4 is twice the sum for
column 3.

The optimum sequence obtained after reordering the events on the

basis of their ranking numbers in column 5 is identical to the optimum
sequences previously given in Tables 5.5 and 5.9. The optimum sequence
obtained in column 6 is the same except that event 3 comes below event 2
because it has a lower ranking number. It will be seen in the next chapter
that the ranking numbers in columns 5 and 6 of Table 5.12 are very close
to the cumulative RASC distances resulting from scaling. There is a
natural transition from ranking to scaling as also pointed out by Kemple
et al. (1990).

The preceding method of averaging probabilities is a method of

probabilistic ranking which is equivalent t o a method described by
Kendall (1975, p. 151). The method was used for ranking by Blank and
Ellis (1982, p. 418) along with a slightly different method to synthesize
local range data found among a group of geological sections (Fig. 5.4). The
modified average probability values for taxa computed by Blank and Ellis
are the same as the ranking numbers of column 6 in Table 5.12, except
that a frequency Fi, was replaced by Fji if Fji >FG.These modified average
probability values cannot be used for ranking or scaling because, on the
average, they first decrease from being close to unity near the top t o nearly
0.5 in the middle of the composite range chart. Next, continuing t o move
in the stratigraphically downward direction, they increase t o nearly 1.O
toward the bottom of this range chart. Blank and Ellis (1982) found that
these modified average probabilities were useful indicators for taxa with
mutually inconsistent local range zones. Suppose that the top (highest
occurrence) or base (lowest occurrence) of a taxon occupies random position
with respect t o the tops and bases of other taxa in the sections. The Blank-
Ellis average probability of such a random event then would be close to 0.5
(its expected value is slightly grater than 0.5 if tops and bases of the taxa
both occur in one or more sections, because the top of a taxon comes above
its base). By successively deleting events with the smaller values, Blank
164

351

25

-
D

al
-

v)
I
C
>
al
W
5 15
L
0,
n
$
Z

!/Threshold
6 7 8 85 9 1
Average nlN

Pig. 5.4 Method of ranking used by Blank and Ellis (1982). Left side: The design of the matrix used to
synthesize local range data found among a group of geological sections. All taxa range endpoints a r e
identified as being a top or base and a r e listed a t the left and across the top of the matrix. The matrix
elements a r e the ratios d N , and contain the empirical stratigraphic positionings of all endpoints found
for a region, taken two a t a time. For example, n2lN2 is the second matrix element and shows that the
Top of taxon A and the Top of Taxon B a r e found stratigraphically separated in N z sections, and the Top
of A is found above the Top of B, n2 times. A row represents a n endpoint's total stratigraphic positioning
compared to all other endpoints with which i t shows a preferred sequence, dN>i. Conversely, d N < b
also shows a preferred (reversed) stratigraphic sequence and was included in the row total as I-nlN. A s
the total for a row approaches +, an endpoint shows a more random stratigraphic positioning, and is not
useful in determining biostratigraphic sequence trends. The threshold a t which a n endpoint is
considered randomly distributed with respect to another or with respect to all endpoints with which it is
physically associated depends on the level of confidence one is willing to accept. Right side: Threshold
value determined for the North Atlantic Ocean database of Blank and Ellis (1982). The horizontal axis
represents the average dN for a taxon as compared to all other taxa with which it occurs. The vertical
axis represents the taxa remaining in the database after successively deleting taxa that fall below a
certain value. The relationship defined for the North Atlantic Ocean database in the main body of the
figure reveals that a t threshold value 0.85, the database maintains a minimum level of confidence and a
maximum number of taxa for further analysis The implication is that taxa falling below the threshold
values are less useful in biostratigraphic classification based on sequential similarities (from Blank and
Ellis, 1982).
 165

and Ellis determined a threshold value of 0.85 for their very large
database of DSDP data (see Fig. 5.4B).

This method must be used with caution because its automated

application could result in the rejection of events from the middle of the
range about where all events (random and nonrandom) have modified
average probability values close t o 0.5. Thus other factors should be
considered as well when this method is applied.

5.6 Conservative ranking methods

As discussed in Chapter 2, the observed highest occurrences of taxa

are probably “too low”, and the observed lowest occurrences “too high” in
any section.

It may be assumed that, within a study region containing a group of

sections, each taxon has unknown true first and last fossilized occurrences.
In conservative ranking methods it is attempted to find the relative order
of these true stratigraphic events. Different methods have been developed
by several authors including Shaw (1964), Edwards (1978) and Guex
(1987). A new method for conservative ranking will be introduced later in
this book (modified RASC, Chapter 8). Most of these methods use observed
positions of events within the sedimentary sequences of the sections .as
well as their relative order. The conservative ranking method introduced
by Rubel (1978) will be used here as an example to illustrate the principles
of this approach labelled as “deterministic” by Guex and Davaud (1984)
and Rubel and Pak (1984). A comparison with the probabilistic ranking
approach also will be made.

Comparison to Rubel’s method

Rubel (1978) has proposed the following method: Suppose that, in a

stratigraphic section, 12 taxa (numbered 1-12) were observed in 5
consecutive samples. The local ranges of these taxa can be represented as
follows:
166

9 10 11 12

5 6 7 8 9 10 11

3 4 5 6 7 8 9 10

2 3 5 6 7 9 10

1 3 5 9

In this tabulation, the taxa are arranged in the order of their

disappearance. Table 5.13 is the corresponding matrix of stratigraphic
relations between the 12 taxa. Each + in Table 5.13 indicates that the
local range of the taxon in the row containing this + is above the
corresponding taxon in the column. The counterpart of + is - signifying
that the first taxon is below the second taxon. Overlap of local ranges is
shown as 0. The three columns in Table 5.13 are for frequencies of , 0 +
and - per row. These row tables are written as a, b and c , respectively.
They can be used for ordering the taxa. For example, ordering the taxa on
the basis of the statistic a is equivalent t o arranging them in the order of
their disappearance. If successive taxa have equal values of a , then they
are ordered according t o their -c values.
Table 5.13 resembles the A-matrix for probabilistic ranking (cf. Table
5.8) of stratigraphic events. However, the A-matrix corresponding t o
Table 5.13 becomes four times as large if highest and lowest occurrences of
all taxa are considered separately as in Table 5.14. Each + in Table 5.13
is equivalent a square block of 4 ones in Table 5.14. Likewise, - becomes a
block of 4 zeros. A zero in Table 5.13 is changed into one of 16 possible
square blocks with its 4 positions occupied by 1, h( =+) or 0 in Table 5.14.
This indicates that Table 5.14 contains more stratigraphic information
than Table 5.13. Figure 5.5 shows all these possible configurations
together with the relations between the ranges of the taxa they represent.
Harper’s (1981) eleven possible relative age relations between two taxa
(see Fig. 2.5) are all represented. In Table 5.14 and Figure 5.5, there are 6
additional configurations because a separation is made between
coexistence of taxa in one or more consecutive samples. Rubel’s (1978)
example has all possible relations between taxa except the situation (not
shown in Fig. 5.5)that two taxa would both occur in one sample only.
 167

TABLE 5.13

Rubel’s matrix of stratigraphic relations between 12 taxa in single section (example of local ranges
+
discussed in text). The row totals a. b and c a r e for , 0 and -,respectively.

1 2 3 4 5 6 7 8 9 1 0 1 1 1 2 a b c
t x + O + O + + + O + + + 8 3 0
2 - x 0 + 0 0 0 + 0 0 + + 4 6 1
3 0 0 x 0 0 0 0 0 0 0 + + 2 9 O
4 - - 0 x 0 0 0 0 0 0 + + 2 7 2
5 0 0 0 0 x 0 0 0 0 0 0 + 1 1 0 0
6 - 0 0 0 0 x 0 0 0 0 0 + I 9 1
7 - 0 0 0 0 0 x 0 0 0 0 + I 9 1
8 - - 0 0 0 0 0 x 0 0 0 + I 8 2
9 o o o o o o o o x 0 0 0 0 1 1 0
1 0 - 0 0 0 0 0 0 0 0 x 0 0 0 1 0 1
1 1 - - - - 0 0 0 0 0 0 x 0 0 7 4
12 ~ . . . . . . - 0 0 0 x O 3 8

Suppose that local ranges for the taxa are available for another
section. A table similar to Table 5.13 then can be constructed for this other
section. The tables for the two sections can be superimposed on one
another and combined into a single new table using the following algebra
(Rubel, 1978, p. 244): & = + + +,
-&-=-, = & O = O and -&=O. I t is
+
implied that O& = 0 and O&-= 0. If one or both taxa are missing in one of
the sections, the matrix element ( + ,- or 0) for their relation in this section
is unknown. Writing x for such a n unknown element, the following
+
combinations can be added: &x = ,-&x =-, O&x = 0 and x&x =x. +
It is possible t o add more sections to a combination of two sections.
The matrix resulting from adding all available sections for a region is
independent of the order in which the sections are added to one another.
+
A in this final matrix, means that, of the two taxa compared, one occurs
above the other in all sections considered. The is accompanied by a - as +
its counterpart. A zero means that the two taxa coexisted in at least one
sample in at least one section. Great importance is given to coexistences
of taxa because the ranges in the composite standard are extended to cover
all observed coexistences of taxa. Obviously, this makes conservative
ranking methods sensitive to reworking and stratigraphic leaks. Such
effects should be eliminated before application of the method.
168

TABLE 5.14

A-matrix for Rubel’s example of 12 local ranges. Each taxon was assigned separate code numbers for its
lowest and highest occurrence, respectively. See text for further explanation.

1 2 3 4 5 6 7 8 9 10 11 12
I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 I 8 19 20 21 22 23 24 A,
l x l l l h l l l h l l l 1 1 1 1 h l 1 1 1 1 1 1 21.5
2 0 ~ 1 1 h l l l h l l l1 1 1 1 h l 1 1 1 1 1 1 20.5
3 0 0 ~ 1 0 1 1 1 0 1 h lh l 1 1 0 1 h 1 1 1 1 1 165
4 0 0 0 x O 1 1 1 0 1 h l h l 1 1 0 1 h l 1 1 1 1 15.5
5 h h l l x l l l h l l l 1 1 1 1 1 1 1 1 1 1 1 1 21.5
6 0 0 0 0 0 x h h 0 1 0 1 0 1 h l 0 1 0 1 I I 1 1 115
7 0 l ) 0 0 0 1 x 1 0 1 0 1 0 1 h l 0 1 0 1 I 1 1 1 125
8 0 U 0 U h h 0 x 0 1 0 1 0 1 h l 0 1 0 1 1 1 1 1 115
S h h l l h l l l x l l l 1 1 1 1 h l 1 1 1 1 1 1 210
~ ~ ~ ~ ~ ~ 1 n I ~ OO h 0 O 0 h 00 0
1 ~0 10 hh l 1 1 7.0
I I i l O h h O 1 1 I 0 I x 1 h I I I 0 1 h l 1 1 1 1 16.0
I ~ I I I ~ I I ~ I I I I I O ~ O C h ~ OI h~ X0 1 0 1 0 1 1 1 65
I : 1 0 I l h h O 1 1 1 0 1 h 1 X I I 1 0 1 h l 1 1 1 1 16.0
l l O U U O 0 0 0 0 0 h O h O x O h 0 1 0 1 h l 1 1 7.0
1 5 O U O O ~ h h h O 1 0 10 1 X I 0 1 0 1 I 1 1 1 11.5
I i i 0 0 0 0 0 0 0 0 0 h 0 h O h O x 0 1 0 1 h l 1 1 7.0
1 7 h h I I 0 I I I h 1 1 1 1 1 1 1 X I I 1 1 1 I 1 205
1 B 0 I I l J i l O l l 0 0 O ~ O 0 0 0 0 0 O x O h O h h h 20
1 9 0 0 I 1 I 1 0 1 1 I 0 I h I h l 1 1 0 1 X I 1 1 1 1 100
2 0 0 0 0 0 0 u 0 0 0 0 0 0 0 0 0 0 O h O x O h h h 20
2 1 0 0 0 0 0 0 0 0 0 h 0 1 O h O h 0 1 0 1 X I I I 75
2 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 O h O h O x h h 20
2 J 0 U l J 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 1 X I 40
2 . i 0 I I 0 0 0 U 0 0 I ~ 0 0 0 0 0 0 0 h h h h h h O x 30

In terms of graph
~- theory, Table 5.13 is the adjacency matrix for a local
range chart represented as an interval graph. However, after addition of
one or more other sections, using the preceding algebra, it may not be
A B C D E F C H I
3 4 5 c. 5 6 11 6 7

HlO(H11) H7(H2)

I:- I::I I-
:I - 3
Pig. 5.5 Graphical representation of all possible configurations of relations between the local ranges of
two taxa in Rubel’s (19781 example. Numbers of taxa used for example a r e same a s in Tables 5 13 and
5.14. Each relation corresponds to a square block of four numbers (1, h = 0.5 or 0) in the upper triangle of
‘Fable 5.14 and its counterpart in the lower triangle. All Harper’s (1981) possible relative age relations
between two taxa (H1 to H 1 1 with numbers a s in Fig. 2.5) a r e represented.
 169

It

Fig. 5.6 Rubel’s (1978)possible explanations of potential inconsistencies for superpositional relations of
3 events in 3 or more sections. In both spatial distribution patterns (A and B), coexistence of the taxa ( a l ,
a2 and ag) cannot be observed in any of the sections (Sl,S2 and S3).

possible to directly represent the resulting table as a range chart because

it may contain inconsistencies preventing its representation as an interval
graph. Figure 5.6 (from Rubel, 1978) shows two inconsistencies of this
type. Rubel (1978) would accept such inconsistencies as real phenomena
only if their existence is reconfirmed by similar contradictory
superpositional relations in other sets of three sections. Unusual
superpositional relations in three sections as shown in Figure 5.6 normally
will not be preserved in the final table if the latter is based on many
sections with other types of superpositional relations for the same three
events. It is noted that combining sections by means of the probabilistic
ranking method results in an optimum sequence (e.g. Table 5.14) that can
be represented a s a range chart in which the highest and lowest
occurrences of each taxon have average positions with respect to those of
all other taxa. As already pointed out in Chapter 2, if the ranges of the
taxa in a range chart of this type are plotted along a geological time scale,
they are shorter than those in range charts based on conservative ranking
methods. This is because superpositional relations with scores less than
0.5 are ignored in probabilistic ranking by setting them equal to zero.
170

5.7 Three-event cycles

Worsley and Jorgens (1977) have found that the algorithm of Section
5.3 does not necessarily yield an optimum sequence because cyclical
inconsistencies may occur in which more than two events are involved.
Their original example of cycling events is shown as the first matrix of
Table 5.15. When the algorithm is applied, the original S-matrix reoccurs
after every set of six consecutive iterations. Hence an optimum sequence
could never be determined by means of the preceding algorithm.

In the example of Table5.15, A occurs more frequently before B

(SAB > SBA),B before C (SBC> SCB), and C before A (SCA> SAC).The
three events A, B and C are involved in a cyclical inconsistency and are
said t o form a three-event cycle. It is useful t o represent this type of
situation by means of a graph. The relationships of Table5.15 are
represented by arrows in the graph shown in Figure 5.7. The three-event
cycle involving A, B and C is immediately apparent in Figure 5.7 because
the arrows in the triangle ABC point in the same direction at both sides of
each of the vertices of this triangle.

If there are no cycles, all inconsistencies can be eliminated by

disregarding situations in which SQ < Sji. Suppose that each situation
+
SQ2Sji is indicated by a sign for Sij in the upper triangle above the
diagonal of the S-matrix where j > i and a - sign for the corresponding
element in the lower triangle where j < i. Then the S-matrix of
Table 5.4D which is a final order relation matrix would be replaced by a
matrix with exclusively + signs in the upper triangle and - signs in the
lower triangle. If a 3-event cycle occurs, it is not possible to achieve a clear
subdivision of this nature as is illustrated in Figure 5.8 for an artificial
example. The events of Figure 5.8 are indicated by means of letters. C, F
and K form a 3-event cycle. The elements in the first two rows could be
tested by means of the previous algorithm. However, iterations would
continue indefinitely for the elements in the third row which is for one of
the cycling events (C). The event in the margin of the third column of
Figure 5.8 can be scanned by putting a “window” on it in the computer
algorithm. For the 3-event cycle of C, F and K, this window will begin
showing the sequence CKFCKF ... which can be readily detected. Once the
events involved in a cycle have been identified, the + sign corresponding
to the pair of scores with the smallest difference ISg-Sjil can be allowed to
remain in the lower triangle. In the algorithm, this is accomplished by
temporary replacement of its scores by zeros. This replacement is
 171

TABLE 5.15

Example of cycling events (initial matrix from Worsley and Jorgens, 1977). Unlike the example of Table
5.4, the algorithm for ordering does not yield a n optimum sequence because the initial matrix returns
after 6 iterations. Note that event D does not participate in the cycling.

x 232 x 243 x511 x 322 x 423 x 151 x 232

1 xs1 5 x 11 2 x43 4 x23 3 x22 2 x32 1 xs1
42 x 3 32 x 2 23 x 2 15 x 1 51 x 1 24x 3 42x 3
074 x 470 x 740 x 047 x 407 x 704 x 074 x

Fig. 5.7 Three-event cycle (ABC) in set of four events is characterized by successive arrows pointing in
same direction a t both sides of vertices (A, B and C). Arrow between two events indicates that one event
precedes other event.

temporary if ranking will be followed by scaling because for scaling,

elements in the lower triangle may be larger than their counterparts in
the upper triangle.

It is possible that two pairs of scores for events involved in a 3-event

cycle have equal smallest difference values, or that all three pairs have
equal differences. In those situations only the first pair encountered will
be ignored. An example is provided in Table 5.16.

For this example, the data of Table4.10 were run setting the
threshold parameters equal to h, = 7 and m,l = 5, respectively. For n = 26
events , it is possible to make n(rt-1)/2=325 comparisons. However,
because of the treshold m,l=5, forty pairs were not used. The presorting
option was used (see Table5.11) and the 26 events were reordered by
172

means of the modified Hay method using the ranks in the last column of
Table 5.11. The final result is shown in Table 5.17. A three-event cycle
involving events 25, 27 and 69 was identified with the corresponding
output shown in Table 5.16. The event positions printed below the cycling
events are temporary and can be used to identify which pair of events (11
and 12) was ignored in order to break the cycle.

In the original input, the three cycling events were encountered

together in four wells: Freydis (69, -27,25), Gudrid (69,25,27),Bonavista
(25,27,69)and Dominion (27,25,69). In these expressions, relative order
is indicated by means of a comma and coeval events are separated by a
comma followed by a hyphen (e.g. in Freydis, 69 and 27 are coeval and both
precede 25). For abbreviation, the four expressions can be rewritten as
(2-31,213,132,312) where 25,69 and 27 have been replaced by 1 , 2 and 3,
respectively. Two of the three events were encountered together in seven
wells with relative orders (21, 21, 13, 12, 21, 13, 32). The scores of
Table 5.16 can be obtained by counting subsequences for two events
(e.g.21 occurs 5 times while 12 occurs 3 times). All t h r e e events

A B O D E @ G * . . @ L.**

+ + + + + *.* + + *..

\
Fig 5 . 8 Graphical illustration of algorithm developed to locate three-event cycle. Elements in
successive rows of upper triangle a r e tested proceeding from left to right. Row and column interchanges
only take place when element is less than its counterpart in lower triangle. In example, element circled
in margin C will be replaced by K which, in turn, will be followed by F. Cycle C K F will repeat
indefinitely.
 173

TABLE: 5.16

Selected output from RASC program including information on a single 3-event cycle encountered when
data of Table 4.10 a r e run with h, = 7 and m,l= 5. See text for explanations.

RUN FOR 7 OR MORE OCCURRENCES AND 5 OR MORE P A I R S .

C Y C L I N G EVENTS: 27 25 69

EVENT P O S I T I O N S : 11 13 12

MATRIX ELEMENTS :

0.0 2.0 3.5

4.0 0.0 3.0

1.5 5.0 0.0

C(11, 13) AND C ( 1 3 , 1 1 ) ZEROED

RANKING S O L U T I O N O B T A I N E D W I T H :

1 0 2 I T E R A T I O N S O U T OF MAXIMUM 9000

TOLERANCE OF 0.0

participate in a cycle because the preferred subsequences 21, 13 and 32

cannot hold true simultaneously.

In this application, the optimum sequence (Table 5.17) is almost equal

to the result obtained by means of the presorting option (Table 5.11). In
addition to a change in order corresponding to the 3-event cycle, only the
events with ranks 2 1 and 22 have changed places in the sequence. Every
cycle is allowed t o run 100 times before it is broken. Hence the total
number of iterations is 102 instead of 2 in Table 5.16. Extra iterations
may be needed to eliminate possible pseudo-cycles which can develop
initially before a truly periodic cycle appears. This subject will be
explained in the next section which also contains a discussion of the
situations in which cycles involving more than three events can develop.

Cycles tend to occur frequently if one or both of the following two

conditions are satisfied: (1)many small samples are used (e.g. R , < 3),
and (2) the expected values of many of the frequencies P , =S,IR, are close
t o 0.5. The tolerance parameter (b,) can be used in the RASC program to
reduce the number of cycles. If b, is set equal to a positive value (e.g. 0.5 or
l.O), scores with S, +
b, > SJl > Sij will be allowed to occur in the lower
triangle (j< i) in addition to the values SJL< S,. By leaving a certain
174

TABLE 5.17

RASC program output of optimum sequence ofdata of Table 4.10with k,=7 and m,l= 5.

Sequence Fossil Range Fossil

Position Number Name

1 17 0- 2 Asterigerina gurichi
2 16 1- 3 Ceratobulimina contraria
3 67 2- 4 Scaphopod s p l
4 18 3- 6 Spiroplectammina carinata
5 21 3- 6 Guttulina problema
6 20 5- 7 Gyroidina girardana
7 15 6- 8 Globigerina praebulloides
8 26 7-10 llvigerina dumblei
9 70 7-12 Alabamina wolterstorffi
in 24 8-1 I Turrilina alsatica
11 27 10-12 Eponides umbonatus
12 69 11-13 Nodosaria s p 8
13 25 12-14 Coarse arenaceous spp.
14 31 13-16 Pteropod s p l
15 29 13-16 Cyclammina amplectens
I6 34 15-17 Marginulina decotata
17 41 16-18 Plectofrondicularia spl
18 42 17-19 Cibicidoides alleni
19 30 18-20 Cibicidoides blanpiedi
20 36 19-23 Pseudohastigerina wilcoxensis
21 45 19-22 Bulimina trigonalis
22 57 21-23 Spiroplectammina spectabilis
23 46 22-25 Megaspore spl
24 50 22-25 Subbotina patagonica
25 54 24-26 Textularia plummerae
26 56 25-27 Glomospira corona

amount of “noise” in the system, an optimum sequence then is obtained

more rapidly requiring less computing time.

5.8 Higher-order cycles and pseudo-cycles

Suppose that four events (A, B, C and D) with Sij=Sji (i=A,B,C,D;

j=A,B,C,D; i * j ) are subject to the relationships SAB> SBA,SBC> SCB,
SCD> SDCand SDA> SAD.
This situation was in fact shown in Table 5.15. Worsley and
Jorgens (1977) assumed t h a t all four events participated in t h e
inconsistency. However, when the algorithm of this paper is applied, only
the events A, B and C are involved in what is called a 3-event cycle.

In general, it can be shown that, if S,j.=Sji Citj)for four events, then

there must be two 3-event cycles in the system for the situation defined a t
the beginning of this section. The scores for A in comparison to C satisfy
 175

either SAC> SCAor SCA> SAC. If SAC> SCA,A, C and D form a

3-event cycle; if SCA> SAC,A, B and C form a cycle. Likewise, either A, B
and D or B, C and D form a 3-event cycle. If the algorithm is applied, a
3-event cycle (and not a 4-event cycle) will be identified (cf. Table 5.16).
When this cycle is broken, the other cycle either remains in the system and
would be identified next, or it is broken at the same time as the first cycle.
Whether or not two cycles will be identified depends on the relative
magnitudes of the differences ISQ- Sjil.
A true 4-event cycle with SAB>SBA, SBC> SCB,SCD> SDC,and
SDA> SADarises only if SAC=SCA and SBD=SDB as illustrated in
Figure 5.9. Higher-order cycles including the 5-event and 6-event cycles
which also are shown in Figure 5.9 only occur if all arrows for arcs on the
circumference of the graph point in the same direction while all indirect
connections between vertices are undirected with Sij=Sji ( i z j ;j z i + 1).
Higher-order cycles are identified and eliminated in the same
manner as 3-event cycles. It is noted that in Gradstein and Agterberg
(1982) all pairs of scores with equal minimum differences were ignored
whereas, in the algorithm described here, only the first pair encountered
will be ignored. Four-event cycles frequently occur in practice but 5-event
cycles are rare. In numerous runs of RASC I have encountered a 6-cycle
only twice. The RASC program would identify and break cycles of up to
nine events. The problem of dealing with cycles of several stratigraphic
events also has been discussed by Salin (1989).
The concept of a pseudo-cycle is illustrated in Figure 5.10. The initial
order ABCD is changed into ACDB after four iterations. The sequence
ACDB contains a single 3-event cycle (ACD) and reappears with a
periodicity of six iterations. When a window is placed on the first event,
the observed sequence is ADCBADCADCA ... This initially would suggest
a 4-event cycle involving all four events. However, this pseudo-cycle is
unstable and is automatically replaced by the 3-event cycle for A, C and D.

5.9 The influence of coeval events

In Hay's original method, coeval events are ignored. On the other
hand, Davaud a n d Guex (1978) a n d Rube1 (1978) in their methods
assigned more weight to ties (coeval events) than is done in the modified
Hay method. In Section 5.3 the practice of several authors including
176
B

E D

Fig. 5.9 Cycles of more than three events can occur when all events, except those involved in cycle, a r e
pairwise simultaneous (relative frequency P , is equal to 0.5). Pair of events that a r e coeval on average
have connecting lines without arrows in examples for 4-, 5- and 6-event cycles shown.

.Ancn ~ B C A E B D A
A x + + - L)x 0 - + cx - t -
B - x + o B o x + - B + x o -
c-- x t c + - x - D- 0 x t

D + o - x A - + t Y A + + - x

BCDA ACDB CAB

B x + o - A X t - + Dx- to
c- x + - c- x + - c +x - -
D o - x t D+- x o A - + x +
A + + - x R - + o * R n + - x

PDAB ADCB AACB

cx t - - A Y - t t D x t - 0
n- x t 0 D * x - 0 A - T + I
A + - x + c- + x - c +- x -
D + o - x R - 0 + x R o - r x

~ A D B ACDB
c x - + - Ax + - +
A + x - + c- x + -
D - t x o D +- Y 0
B t - o x D - + o x

Fig. 5.10 Illustration of pseudo-cycle (ADCB) which initially develops when the algorithm is applied but
is automatically replaced by the three-event cycle (ADC). Events with hats a r e being observed a t a
“window” and checked for periodicity in the algorithm.
 177

TABLE 5.18

KASC program output of optimum sequence for Hay example after modifications of SEQ file of Table 5.3
(cf. Table 4.6). A. Additional information for Paleocene was used. B. Guex levels were used for data
reduction.

A Sequence Uncertainty Event Event

Number Range Code Name

1 0-3 9 HI Discoaster tribrachiatus

2 0-3 10 LO Discolithus distinctus
3 2-5 6 LO Hhabdosphaera scabrosa
4 2-6 8 LO Discoaster cruciformis
5 3-6 4 LO Coccolithus solitus
6 5-7 7 LO Discoaster minimus
7 6-8 3 1'0 Coccolithus germanicus
8 7-9 1 LO Discoaster distinctus
9 8-10 5 LO Discoaster gammation
10 9-11 2 LO Coccolithus cribellurn

B Sequence Uncertainty Event Event

Number Range Code Name

1 0-2 10 LO Discolithus distinctus

2 1-3 9 1-11 Discoaster tribrachiafus
3 2-5 8 LO Discoaster cruciforrnis
4 2-6 6 LO Rhabdosphaera scabrosa
5 3-8 7 LO Oiscoaster minimus
6 4-7 4 LO Coccolithus solitus
7 6-8 5 LO Coccolithus gammation
8 7-10 1 LO Discoaster distinctus
9 7-1 1 3 LO Discoaster germanicus
10 8-11 2 1.0 Coccolithus cribellurn

Kendall(1975), and Brunk (1960) who scored ties as 0.5 above and below
the principal diagonal of the matrix for frequencies. However, arguments
that ties should be ignored in some situations have been presented by
Hemelrijk (1952) and Tocher (1950). It has already been pointed out that,
in the absence of cycling (see Section 5.7), the modified Hay method
produces exactly the same optimum sequence as the original Hay method.
178

In the methods of Davaud and Guex (1978) and Rube1 (1978), occurrences
of fossil species are considered to be coeval if they are observed t o the
coeval at least once. For example, even if fossil A is observed to occur
above fossilB in several sections, their coexistence in a single section
results in the two fossils t o co-occur in the standard contructed on the basis
of all sections. Clearly, more weight then is assigned to ties than in either
the Hay method or modified Hay method. Guex and Davaud (1984) have
made extensive use of graph theory in developing their technique. This
allowed them t o construct an optimum sequence of multiple events which
may be subdivided into parts called “Unitary Associations” (see Section
3.5) that can be identified in the original sections and used for correlation.

In Chapter 4 it was pointed out that the results of ranking (and

scaling) depend on how the original data are coded. For the Hay example,
it was noted that scoring ties for coeval events resulted in bias do to
artificial truncation on the stratigraphically lowest levels of some sections.
Several of the nannofossils used in the example already existed before the
Eocene and their entries with respect to one another in the Paleocene were
known for two sections. Use of this information changed the partial SEQ
file for the Media Agua Creek section (see Table 4.6). The optimum
sequence of Table 5.5 is changed into that of Table 5.18A when a revised
SEQ file with data for the Paleocene in the two sections is used. The
revisions in the optimum sequence are minor and restricted t o the lower
part of the optimum sequence.
It also was noted in Chapter 4 that the method of preprocessing by
coding events from maximal horizons (cf. Fig. 4.4)gives another type of
SEQ file (cf. line 2 in Table 4.6).Table 5.18B shows the optimum sequence
obtained for the 10 events of the original Hay example after coding them
from Guex levels for all 9 sections. Again the resulting revisions are
relatively minor. From the discussions in Chapter 4, it may be concluded
that the optimum sequence of Table 5.18A is marginally better than the
one of Table 5.5 whereas that of Table 5.18B would be marginally worse.
However, for this example, it is not possible to prove whether or not minor
revisions of this type are significant. In magnitude they are comparable to
the types of changes that arise when one or more of the threshold
parameters h,, m,l and b, are modified.
 179

CHAPTER 6
SCALING OF BIOSTRATIGRAPHIC EVENTS

6.1 Introduction

The RASC computer program for ranking followed by scaling of

stratigraphic events was originally published with documentation in
Agterberg and Nel (1982a, b). Many examples of scaled optimum
sequences can be found in Gradstein et al. (1985). The purpose of this
chapter is t o review the scaling method in detail using relatively small
datasets. First the principle of scaling is explained by applying it to simple
artificial examples and by approximating the transformation of the
relative frequencies PG into distances 20, as performed in RASC, by a
linear transformation which is easy to understand.
In the artificial examples of Figure 6.1, observed occurrences of two
stratigraphic events (A and B) in 12 sections are compared with one
another. An additional event (C) is considered in Artificial Example 4. As
a rule, biostratigraphic events are observed only in a subset of the total
number of sections ( N )in a study region. In Artificial Example 1, N = 12
but A occurs only in N A = 5 and B in N B = 6 sections. The number of
sections NA,B = 2 with both A and B present is even smaller. In these two
sections, relative stratigraphic position of A is above that of B. This
relation can be quantified by writing NAB = 2 and N B A = 0, where AB
indicates A above B and BA is A below B. In the other examples of
Figure6.1, A-B denotes that A and B were observed to be coeval with
frequency NA-B(e.g. NA-B = 4 in Artificial Example 2).
In total, three threshold parameters have to be set a t the beginning of
a RASC run: h,, m,l and m,2 with h, 1 .m,2? m,l. The critical value k,
indicates that an event will only be used for computing if it occurs in a t
least h, sections. If one would set k, = 6 in Artifical Example 1, the event
A would not be used for ranking and scaling. The parameters m,l and m,2
control minimum number of pairs of events to be used for computing
optimum sequences in ranking (modified Hay method, see Section 5.4) and
scaling, respectively. If m,l = 1and m,2 = 4 in Artificial Example 1(with
h, 2 5 ) , A and B would be compared for ranking but not for scaling. If h,
180

and mC2are increased, statistical precision of results is improved but fewer

events are considered.

The methods of ranking introduced in the previous chapter produce a

simple answer for the examples of Figure 6.1. If NAB > N B A as in
Artificial Examples 1 and 3, the ranking result is AB. The optimum
sequence for the fourth example is ABC, and “undecided” for Artificial
Example2 where a decision cannot be taken. The scaling technique is
conceptionally more complex than ranking. Using the frequencies N A B ,
N B A , N A - B and N A , B , a single relative frequency P A B =
(NAB4- 0.5NA-B)/NA,B is computed. Obviously, PBA = 1 -PAB. The
principle of scaling is that the frequency for inconsistencies PAB is
transformed into ZAB = @ ‘ - ~ ( P A Bbeing
) an estimate of the interval
between mean positions of A and B along a distance scale (RASC scale). @
represents fractile of the normal distribution in standard form. If it is
found that PAB = 1 for the situation that A and B are relatively close
along the RASC scale, PAB = 1 is replaced by a probability which is less
than 1 and the corresponding interval is set equal t o ZAB = qc . In
Artificial Example 1,NA,B = 2 with PAB = 1. If this relation would be
used in conjunction with other frequencies (e.g. for “indirect” estimation,
see later), we could choose PAB = 0.90 with qc = 1.282. The “default”
value in RASC is qc = 1.645 for P = 0.95.

The transformation 0-l can be approximated by t h e linear

transformation Z*AB = 2.93 (PAB-0.5) as illustrated in Table 6.1. It is
useful to define an interval 2 = Z* = 0 for P = 0.5 when one is not able to
decide whether A should be above or below B in the optimum sequence as
in Artificial Example 2. In Artificial Example 3, PAB = 5/8 which yields
ZAB = 0.319 and Z*AB = 0.366. In Artificial Example 4, PAB = 3.5/5
which is slightly greater than 5/8 in Example 3. The resulting distance
Z*AB = 0.59 (ZAB = 0.52) also is slightly greater.
For Example4, PAC = 5/6 with Z*AC = 0.98 (ZAC = 0.97), and
PBC = 7/9 with Z*BC = 0.59 (ZBC = 0.77). These three estimates of
distance are not mutually consistent. For example, Z*AB.C= Z*AC -
Z*BC = 0.29 provides an indirect estimate of the distance between A and B
which differs considerably from the direct estimate Z*AB = 0.59. This
type of inconsistency can be ascribed t o small sample sizes and can be
+
eliminated by averaging ; e.g. Z*AB = 0.5 (Z*AB Z*AB.C) = 0.38 which
is close t o ZAB = 0.36. Especially when there are many indirect distance
estimates, such averages are more precise than direct distance estimates.
 181

Artificial Example 1

Artificial Example 2

Artificial Example 3

Artificial Example 4

Fig 6.1 Graphical illustration of RASC method for ranking and scaling of stratigraphic events in many
stratigraphic sections (shown a s vertical lines). Ranking in the stratigraphically downward direction
provides optimum sequences AB (A stratigraphically above B) in Examples 1 and 3,A-B (undecided) in
Example 2, and ABC in Example 4. Scaling gives distance estimates of intervals between successive
events along a linear (RASC) scale. The distance between A and B is estimated a s (1) 1.28, (2) 0.00,
(3) 0.32 and (4) 0.36 for Artificial Examples 1,2,3and 4,respectively (from Gradstein e t al., 1990).

In RASC, the averaging process is refined by considering differences in

sample size. For example, P = 1.514 for N = 4 is less precise than
P =4.5/12 for N = 12 although their 2-values are the same. The second Z -
value is given more weight in the calculations because it is based on a
larger sample (see Section 6.2).

The linear transformation was introduced here t o illustrate the

concept of scaling. In practice, it is better to use the normal distribution as
in RASC. This is because a linear transformation would imply that the
182

TABLE 6.1

Example of Z-values for selected relative frequencies P . The Z*-values in last column are linearly
related to the frequencies and are approximate Z-values.

P z Z*
0 00 -Pc -2.930

0 05 -1.645 -1.319

0 10 -1.282 -1.172

0 20 -0.842 -0.879

0 30 -0.524 -0.586

0 40 -0.253 -0.293

0 50 0.000 0.000

0 60 0.253 0.293

0 70 0.524 0.586

0 80 0.842 0.879

0 90 1.282 1.172

0 95 1.645 1.319

100 4c 2.930

frequency density function of the interval between two events along the
RASC scale is uniform. This, in turn, would mean that frequency density
functions of individual events along the RASC scale would have different
shapes depending on the value of Z*; e.g. for Z*AB = 0, A and B would
have U-shaped density functions with local minima a t their mean
locations. It is more realistic t o assume that the individual species have
density functions with maxima a t or near their mean values. The mode
and mean coincide for the normal (Gaussian ) curve model used in RASC.
This model is not satisfactory for small densities in the tails where
artificial truncation is applied when the cumulative frequency of the
sample is observed t o be either 0 or 1 (see before). It is good to keep in
mind that decrease in density away from the mode could be different for
different taxa. Also, for the same species it could be different in the
stratigraphically upward and downward directions (cf. Chapters 2 and 9).
The scaling algorithms presented in this chapter form the second part
of the RASC program for ranking and scaling of biostratigraphic events
and other events which can be uniquely identified. An optimum sequence
constructed by means of a ranking algorithm provides the starting point
 183

for estimating average “distances” between successive events. The

frequency of cross-over (mismatch) of the events in the sections is used for
this purpose. These distances are clustered by constructing a dendrogram
which can be used as a standard and permits definition of average interval
zones (cf. Fig. 2.2). This chapter will include artificial examples in which
the theory of scaling is illustrated and tested by applying it to sets of
random normal numbers in computer simulation experiments.

6.2 Scaling versus ranking

The techniques described in this chapter have in common t h a t

distances are estimated between successive events in the optimum
sequence obtained by the ranking algorithms described in the previous
chapter. In a ranking, the successive events follow each other and no
allowance can be made for the situation that some events should be closer
together than others along a relative time scale. It can be useful t o
position the events along a scale with variable intervals between them.
For example, suppose that two microfossils have observed extinction
points (A and B) in 10 sections with A occurring 5 times above B, and 5
times below B. If a fence diagram were constructed, in which each event is
connected to itself in other sections, the lines connecting event A would
cross those connecting the event B in a number of places. It could be said
that the relative cross-over (mismatch) frequency is PAB = 0.5 because the
number of matches is equal to the number of mismatches. This analogy
generally does not hold true if P is a positive number not equal to 0.5
because, in general, the frequency of cross-overs is partly determined by
the spatial pattern of the geographic locations of the sections. However, if
the number of sections is not too small, the frequency PABalways can be
regarded as an estimate of the probability that A occurs above B. The
interval between A and B along the relative time scale used for scaling
should be nearly zero if PABis close to 0.5, and greater if PABtends t o zero
or one. Suppose that A occurs, for example, 9 times above B and only once
below B. Then A and B should be separated by a longer distance along the
relative time scale, corresponding to PAB= 0.9.

The purpose of the scaling techniques is t o estimate distances in time

between successive events, not only from the cross-over frequencies
between successive events, but also by using the cross-over frequencies
184

between all events with mismatch in location in the observed sequences

for segments of the optimum sequence.

Figure 6.2 from Agterberg and Gradstein (1988) provides an example

of output from a scaling algorithm. The number codes of the events (exits
of microfossils) and the microfossil names are shown on the right side.
Each code is followed by the estimated distance from its event t o the event
below it. These distances have been plotted in the horizontal direction
toward the left. They were clustered during a sequence of linking steps.
The two successive events (32 and 29) in the scaled optimum sequence
with the shortest distance (0.0067) between them were linked first. After
scanning the set of unused interfossil distances, single events or clusters of
events were linked pairwise, a t each linking step, by using the shortest
distances between them until the longest interfossil distance (between 20
and 24) was reached. The resulting clusters based on interfossil distances
in time resemble assemblage zones (cf. Section 2.2).

The solution of Figure 6.2 for 54 taxon exits in 21 wells on the

Labrador Shelf and northern Grand Banks shows a number of distinct and
progressively younger clusters. A shading pattern was used to enhance
the stratigraphically most useful parts of individual clusters. In total, 10
preferred RASC zones are shown. These are separated by relatively long
interfossil distances. Several of such intervals between clusters represent
stratigraphic hiatuses (Gradstein et al., 1985). In order t o construct
Figure 6.2, the output of the RASC program listed in Agterberg and
Nel(1982) was combined with a DISSPLA graphics package (copyrighted
in 1975 by Integrated Software System Corporation). A version of this
DISSPLA program called DENO was published by Jackson et al. (1984).
DENO was used t o construct the optimum sequences and dendrograms of
nine data bases in Gradstein et al. (1985, Appendix I). The input.data for
Figure 6.2 were processed by using the modified Hay method with
threshold parameters h , = 7 , rn,l = 2 and m,2= 4 . The optimum
sequence resulting from ranking was used as a starting point for scaling.
It was slightly reordered during the application of the scaling algorithm
(see later). The distances between successive events shown in Figure 6.2
can be added in order to obtain distance of each event from a common
origin coinciding with the first event (No. 4 in Fig. 6.2). The resulting
RASC distances can be related to geological time (in Ma) on the basis of
those events for which the age is relatively well known (see Chapter 9).
 185

Fig. 6.2 Scaled optimum sequence for 21 wells on Labrador Shelf and Grand Ranks (k,=7, r n ,l = 2 ,
r n , ~=4). Dendrogram values along horizontal axis are interfossil distances ( = i n t e r v a l s between
successive exits) also given in numerical form in the vertical direction. Each distance represents
distance between an event and its successor of which the dictionary code number and name are printed
on the next line. The tenfold zonation is representative for the regional Cenozoic stratigraphy There are
eleven unique events, shown with double asterisks. These unique events occurred in fewer than k , = 7
sections so that they were not used for scaling. Their interfossil distances were estimated later, by re-
inserting them into the scaled optimum sequence on the basis of their relative stratigraphic positions
(with respect to events that were used) in the one or more sections containing them. A shading pattern
was used to enhance the stratigraphically most useful parts of the dendrogram. The large distances on
either side of the Eocene, Oligocene and Miocene assemblages are sedimentary cycle boundaries
(cf. Gradstein e t al., 1985, pp. 146-151).
186

Figure 6.3 shows D E N 0 output for the Hay example (cf. Fig. 4.2,
Table 5.5). All 10 events were used and the threshold parameters m,l and
m,2 were set equal to 2. The relatively short intervals between events 1 to
7 in Figure 6.3b reflect the fact that these events tend to be coeval on the
average in the lower parts of the sections (see Fig. 4.2). On the other hand,
events 8,9 and 10 tend to occur above the others. Clearly, the dendrogram
(scaled optimum sequence (Fig. 6.3b)) contains more information than the
optimum sequence (Fig. 6.3a). As another example of this, it may be
considered that events 9 and 10 are coeval on the average according t o
Figure 6.3a. This would imply that there is 50 percent probability that
event 9 occurs above 10. However, in Figure 6.3b, event 9 occurs above 10
with distance of D=0.4354. It will be shown in the next section that the
estimated probability P , corresponding t o D satisfies P , = @(I)).
Consequently, event 9 would occur above 10 with probability Pe=@
(0.4354)=0.67 o r 67 percent which is slightly greater than 50 percent.
Although W (event 9) occurs three times above A (event lo), and h three
times above W in Figure 4.2, it also can be seen that if W occurs above A ,
the latter event is coeval to six (Section B), one (Section G) and two
(Section H) other events, respectively. On the other hand, if A occurs
above W, the latter event is not coeval to any other events. Because all
possible pairwise comparisons are considered simultaneously in scaling,
event 9 (W) is placed above 10 ( A ) in the scaled optimum sequence instead
of at the same position.

6.3 Statistical model for scaling of stratigraphic events

The existence of events which interchange places with one another in
different sections can be explained by assuming t h a t each event is
described by a different probability distribution. As pointed out before, the
exact probability distributions of the events are not known. However, it
can be assumed that the distributions of the direct and indirect distance
estimates are approximately normal because these are averages of two and
three event distances, respectively, and averages tend t o be normally
distributed (cf. Fig. 2.18). It will be shown that this allows estimation of
the parameters of the model. An advantage of this statistical approach is
that, later, the fitted model can be tested against the observed data. This
 187
OPT I M U M F O S S I L SEQUENCE

6
5
R

9 3

1 ,c

br

I- >

INlER~OSSIl DISTANCIS

Fig. 6.3 D E N 0 output for the Hay example (from Agterberg and Gradstein, 1998). The clustering of
events 1 to 6 in the dendrogram (b) reflects the relatively large number of cross-overs and many coeval
events near the base of most sections used (cf. Fig. 4.2).

final testing either verifies or negates the results obtained by means of the
statistical model.
Figure 6.4 shows the basic model initially adopted for the scaling
algorithms. Each event (e.g. A) would assume a position XAi in section i
where X A ~is the distance to A from an origin with arbitrary location along
the relative time scale (x-axis in Fig. 6.4). The distance x ~ isi assumed to
be the realization of a random variable X A whose probability distribution
is shown in Figure 6.4. Similar random variables are defined for the other
events B, C,...
The random variable X A satisfies the normal (Gaussian) probability
distribution N ( E X A , u2) with expected (or mean) value EXA and
variance u2. The mean values of the events differ from one another but the
standard deviations of all events are assumed to be equal to u in the model
of Figure 6.4.
188

Distance ( x ) along relative time scole

Fig. 6.4 Probabilistic model for clustering of biostratigraphic events (A, B, C, ...) along relative time
scale (x-axis). Relative position of event (for example, A) in section or well is random variable ( X A ) which
is distributed normally around average location (EXA)with standard deviation o.

fc

I
0
I
AAE
-
dAB= x B - xA

Fig. 6.5 Direct estimation of distance AAB between events A and B from cross-over frequency P ( D A B<O).
Random variable DAB(=XB-XA)is negative only when order of A and B in section is reverse of order of
EX* and EXB. Variance O f D A B is twice as large as variance 02 of individual events A and B.

The normal distribution curves for events A and B a r e shown

separately in Figure 6.5. Because the time scale is relative, it will not be
possible to estimate u which determines the scale along the x-axis. (In the
RASC program u2 is set equal t o 0.5, see later) However, it is possible to
estimate the ratio A A B / ( u ~ ~for
) the distance between the population
means AAB = EXB - EXA from the relative cross-over frequency PAB.
 189

For this purpose, PAB is considered to provide a good estimate of the

probability P(XB- X A > 0 ) = P(DAB > 0) which satisfies

(6.1)

This formula follows from the fact that the difference DAB = X B - X A has
a normal distribution N(AAB,20') which is shown in the bottom part of
Figure 6.5. The distance between events A and B for a specific section can
be written as dAB = XB- XA. The hatched area in Figure 6.5 is for
P(DAB<O)= ~ - P ( D A B > O ) . If represents fractile of the normal
distribution in standard form, it follows that

(6.2)

Consequently,
P(D > O ) = @(AAB/0d2)
AB (6.3)

Fig. 6.6 Indirect estimation of distance AAB between events A and B from cross-over frequencies with
event C. Indirect distance DAB,C=DAB-DBC has variance which is four times as large as variance of
individual events A. B and C.
190

A precise estimate of PAB which would allow the determination of

AAB is seldom available in practical applications because this would
require a very large number of sections containing both A and B.
However, it generally is possible to estimate AAB indirectly by using pairs
of cross-over frequencies linking A and B to other events; for example, by
using the pair PAC and PBC. A distance of this type will be written as
DAB.C. As illustrated in Figure 6.6, DAB.C= DAC - D B C is normally
distributed with N(AAB,4u2). Because u2 is arbitrary (0determines scale
along x-axis), the variance of the normal distribution was set equal to the
constant u2 = 0.5. As a result of this simplification, it follows that

(6.4)

In the middle term of Equation (6.41, the event C can be replaced by

any other event from which an indirect estimate of AAB can be obtained.
In practice, it usually turns out that there are many events showing
inconsistencies with both events for which the interval A along the x-axis
is being estimated. Averaging of many indirect distance estimates yields a
more precise estimate of A . Once AAB in Equation (6.4) has been
estimated, it can be used t o estimate P ( D A B > O ) . The resulting
“theoretical” probability should be close to PAB. Although, for model
verification, it is not meaningful to make separate comparisons of this
type, it can be useful t o compare many observed and theoretical
probabilities simultaneously by means of a chi-squared test (see Section
6.11).

It should be kept in mind that the model of Figure 6.4 is not

necessarily realistic because it is unlikely that all events would have the
same normal curve with variance equal t o u2 for their exit location
distributions. However, in practice, an estimate of indirect distance such
as DAB.Cis based on two separate distances (DAC and D B C ) and, each of
these two random variables, in turn, is based on two separate distances
( X A , X c and X B , X c ) although X c is used twice. Hence DAB.Cis based on
three random variables ( X A , X B , and X c ) that cannot be estimated
separately. Because of the central-limit theorem of statistical theory,
DAB.Ctends t o be normally distributed even if the frequency curves of
events A, B and C are not normal and have unequal variances (cf.
Fig. 2.18).
 191

Even if random variables for indirect distances such as DAB.Care not

normally distributed with equal variances, then the computation of an
unweighted or weighted average of a number of indirect distance
estimates, almost certainly, will yield a final estimate of A with a normal
distribution because the central limit theorem applies t o this new
averaging process as well. However, although the final distance estimates
may be precise estimates of the expected values (EXA, EXB, EXc, etc. in
Fig. 6.4) of the exit distributions, the corresponding variances U ~ AU, ~ B ,
u 2 c , ... are not necessarily all equal to 0.5. Neither are all exit
distributions necessarily normal. To assume normality with u2 = 0.5 for
all distributions usually provides a crude approximation of the exit
distributions only (see Chapter 8 for further discussion).

Unweighted distances for Hay example

Table 6.2A shows the relative cross-over frequencies Pij=SijIRij for
the Hay example. The order of the events is that of the optimum sequence
shown previously in Table 5.5. The elements in Table 6.2A are identical to
those in Table 5.3A except that two pairs with Rij = 2 were set equal to zero
because the threshold parameter m c 2 = 3 was used. Each of the
frequencies of Table6.2A was changed into a fractile of the standard
normal distribution or Z-value (see Table 6.2B).

Table 6.1 shows Z-values for selected relative frequencies. Because

Pji = 1-PQ, it follows that Zji = -ZQ. When the optimum sequence is
used as a starting point, all or most of the Z-values in the upper triangle of
the Z-matrix are positive. Negative values occur in the upper triangle
only for elements with PQ< 0.5 corresponding to events whose scores were
ignored in order to break a cycle in which these events were participating
during ranking by means of the modified Hay method. It is noted that
scores temporarily ignored for constructing the optimum sequence are
restored to their positions before use of the scaling algorithms of RASC is
initiated.
Clearly, a relative frequency Pij for a small sample will be subject t o
considerable uncertainty and this error is propagated into the Zij-value
derived from it. This is the reason for defining the minimum sample size
mc2 ( = 3 for Table 6.2). It means that Zij-values based on fewer than mc2
pairs of occurrences will not be used. In the original RASC program
(Agterberg and Nel, 1982a, b) no distinction was made between mcl and
192

m,2. However, later work has shown that better results can be obtained by
setting m,2 > m,l. For the example of Table 5.3, mc2=3 and m,l= 1.
When an average distance between two events is estimated from Z-
values for 10 events, it could be based on as many as nine seperate
estimates of the distance. The direct estimate of the distance between
events i and j follows from Z ~ and
J the indirect estimates involving other
events h follow from the differences Zik - Zjk ( h # i j ) where i a n d j = i + 1
are successive rows. However, because Zij = -Zji, the differences
Zkj - Zki ( h z ij),where i and j = i + 1 are successive columns, also can be
used. For example, the direct estimate of distance between events 4 and 7
which occur i n columns 5 and 6, respectively, satisfies D(4-
7 ) = Z56= 0.210. The corresponding i n d i r e c t e s t i m a t e s a r e
z16-z15 = 1.645-1.068 = 0.577, 2 2 6 - 2 2 5 = 1.282-0.524 = 0.758,
and six other, similar differences between Z-values in adjacent columns.
The differences for all pairs of events are shown in Table 6.2C.
In the RASC program, Z-values in the upper triangle are used only.
The lower triangle is used t o retain information on sample sizes. Addition
of indirect and direct estimates yields the sum of the N* separate
estimates. For events 4 and 7, Sum= 1.56 (see Table 6.2C). The average of
all N*=9 estimates of the interval between events 4 and 7 amounts to
Sum/9 = 0.174. This is called an unweighted estimate of distance between
successive events in the output of the RASC program. The complete set of
9 intervals is shown in Table 6.3. The cumulative RASC distance or
distance from the first event (No. 9) is shown in the last column of Table
6.3. Because of missing values (see Table6.2) or pairs of cross-over
frequencies which both are equal t o one (see later), distance estimates may
be based on fewer than N* ( = 9 for the example) pairs of events.
Theoretically, the direct estimate of distance (cf. Fig. 6.5) has half the
variance of the indirect estimates (cf. Fig. 6.6). Thus it should be weighted
twice as heavily. This will be done in weighted distance estimation in
which errors in Pi,. due to small sample sizes also will be considered.

Weighted distance estimates

The relative cross-over frequencies Pi,. are calculated from scores ( S G )
on samples of different sizes (Rq). For this reason, it is preferable t o
compute weighted mean distances Aec in which the weights assigned t o
the direct and indirect estimates of distance are primarily determined by
 193

TABLE 6. 2

Unweighted distance estimation to obtain intervals between successive events along RASC distance
scale for Hay example. A. P-matrix of relative frequencies for the 10 events in order of optimum
sequence. Values excluded because of threshold mzc= 3 a r e shown as 000. B. Z-values corresponding to
P-values. Note t h a t threshold qc is equal to 1.645. C. Values a r e differences between values in
successive columns of Table 6.2B. Zero differences for pairs of q,-values a r e shown as 000 and were not
used. Bottom row shows sums for columns with number of values ( N * )used for obtaining sum.

A 9 10 8 6 4 7 5 1 3 2
9 x 3 0/6 5 015 4 014 6 017 7 011 9 019 8 018 6 0/6 8.018

10 3 016 X 2 513 000 3 515 4 515 5 016 4 515 3 514 4.515

8 0 015 0 513 ‘L 000 3 014 4 515 5 015 5 015 4 014 5.0/5
6 0 014 000 000 X 3 014 I 5/3 3 014 2 513 3 014 3.014
4 1 017 I 515 I o/.I 1 014 X 3 516 4 517 4 516 4 516 3.0/6
7 0 017 0 515 0 515 1 513 2 516 Y 3 5/7 4 0/6 3 515 4.516
5 0 019 106 0 015 I 014 2 5/7 3 517 X 4 518 4 016 5.018

I 0 018 0 515 0 015 0 513 I 516 2 0/6 3 518 x 2 515 5.017

3 0 016 0 514 0 Oi4 1 014 I 516 I 515 2 016 2 515 X 3.016
2 0 0/8 0 515 0 015 1 01.1 3 0/6 I 516 3 0/8 2 017 3 016 X

H 9 10 8 6 4 I 5 1 3 2
9 Y 0000 I645 I645 I068 I645 I615 I645 I645 1645

10 0 000 X 0967 000 0524 I282 0967 I282 I150 1282

8 I645 0 96 7 ‘L 000 0674 1 282 1645 1645 1645 1645
6 I6 4 5 000 000 X 0 674 0 000 0674 0967 0674 0674
4 I068 0 52 4 0674 0674 X 0210 0366 0674 0674 0000
7 I645 I282 I282 0000 0210 Y 0000 0430 0524 0674
1 1615 0 96 7 I615 0674 0366 0 000 X 0 I57 0430 0318
I 1645 1 28 2 1645 0967 0674 0 430 0 157 X 0000 0566

J I645 I 150 I645 0 674 0 671 0 524 0430 0 000 x 0 000

2 1645 I282 I645 0674 0000 -0674 - 0318 -0566 0000 X

C 10 8 6 4 7 5 1 3 2
0000 I615 000 0577 0 5i7 000 000 000 000
Y 0967 000 000 0 758 0315 0315 0 132 0 132
0 678 Y 000 000 0 608 0 I63 0000 0000 0000
000 000 ‘L 0674 Ofii4 0674 0 293 0 293 0000
0544 0150 1lOOl1 \ 0 210 0 156 0308 0000 0674
0 363 0 1)” I2 S2 0210 Y 0000 0430 0094 0 150
0678 0678 0971 0308 03fiR \ 0157 0273 0112
I 0363 0 3 fil Ofii8 0293 0244 0273 Y 0000 0566
3 0495 0496 0971 0000 0 150 0091 0.130 ‘L 0 000
4 0363 0 3 F3 0971 0674 0674 0356 0248 0566 x

SullVV’ 3 9803 05618 4 8716 I 1617 I 5619 I fiOl8 1 6918 0 5118 006/8
194

TABLE 6.3

Unweighted distance analysis of values shown in Table 6.2 continued to obtain RASC distances of
events. The origin of the scale is set a t the first event. Consequently, the distance for event 9 is equal to
zero. Event 10 has distance of 0.435. Event 2 has the largest cumulative RASC distance ( = 2.140).

Events N* Sum Interval Distance

1 9-10 8 3.98 0.935 0.435
2 10-8 8 0.56 0.070 0.506
3 8-6 6 4.87 0.812 1.318
4 6-4 7 1.16 0.166 1.484
5 4-7 9 1.56 0.174 1.658
6 7-5 8 1.60 0.200 1.858
7 5- 1 8 1.69 0.21 1 2.069
8 1-3 8 0.51 0.064 2.132
9 3-2 8 0.06 0.008 2.140

the sizes of the samples used to obtain the 2-values. The weight-corrected
equation for estimating the distance between events i a n d j is:

(6.5)

where the weights wij and w0.k are

(6.6)

In order t o derive these equations, use was made of theory of

weighting coefficients (cf. Bliss, 1935; Fisher and Yates, 1964; Finney,
1971. The weights were derived in the following manner. The observed
proportion Po is assumed to be the realization of a random variable P
which is related t o a standard normal variable 2 such that

(6.7)
 195

where s denotes position along the linear scale used.

The proportion P can be assumed t o originate from a binomial random
variable with expected value E(P) = Pij and variance

where Rij, as before, is the number of times that events i a n d j occurred in

the same section.
It is known that, approximately,

where p and z represent the density functions of P and 2, respectiuely.

These equations can be combined into

(6.10)

Each weight wLjis obtained as

-2
1 RIJe
w = - - -
’I &Z) 21VlJ(1 - P L J )
(6.11)

Weights W 0 . k are obtained by addition of similar variances 02(Z) of

the values Z i k and Z j k . If 20 = g,, the Pij value corresponding to qc is used
together with the original R u value in Equation (6.11).
Table 6.4 shows intervals which are weighted distances ~ ~ +
i 1, (i
i = 1,
..., N-1) estimated for successive events in the optimum sequence. For
example, the weighted distance between events 4 and 7 is calculated as
follows. From Table 6.2 it follows, for events 4 and 7 , that R,, = 6,
P,, = 3.5/6 and Z, = 0.210. Consequently, w56= 3.76 (Eq. 6.11 or 6.6).
Likewise, for the same example, w15 = 2.91 and w l , = 1.57. Hence,
w , , , ~= 1.02 (Eq. 6.6). The sum of 9 weights is W = 3.76+1.02+0.8=
15.0 (see Table 6.4). The corresponding sum (numerator, right side of Eq.
6.5) is 2.34. The weighted distance between events 4 and 7 therefore is
196

TABLE 6.4

Weighted distance analysis of values shown in Table 6.2. The Z-values were weighted according to
sample size (see Eq. 6.5 and 6.6 in text). Standard deviations were computed by using Eq. 6.13. Note that
the interval between events 3 and 2 (on bottom row) is negative. As a result, event 9 has RASC distance
(=2.149) whichisless than thatofevent 8(=2.155).

Events W Sum Interval s(i) Distance

1 9-10 10.3 3.27 0.317 0.100 0.317

2 10-8 7.0 1.24 0.176 0.289 0.493

3 8-6 4.7 3.62 0.770 0.203 1.262

4 6-4 9.2 2.44 0.266 0.163 1.529

5 4-7 15.0 2.34 0.157 0.153 1.686

6 7-5 14.8 2.32 0.157 0.085 1.893

7 5- 1 15.2 2.96 0.195 0.082 2.038

8 1-3 12.6 1.47 0.117 0.090 2.155

9 3-2 13.3 -0.08 -0.006 0.124 2.149

Ae = 2.34/15.0=0.157. This value is among the intervals listed in

Table 6.4.

For simplification, Equation (6.5)can be rewritten as:

(6.12)

with

'
N
x = AAB; W = 2 wi
1=1

and
x , = Z A B , w 1 = w AB

x2 = zAC-ZBc' w 2 = w AB.C

with similar expressions for xi ( i = 4 , 5, ...). In these expressions, A and B

denote two successive events, and other events are written as C, D, ... The
 197

weight W and sum Ewjxj for the Hay example were given in Table6.4.
The corresponding standard deviation s(2) shown in the last column of
Table 6.4 is the positive square root of
N'

(6.13)

As before, the number of pairs of 2-values used for estimation is written as

N*. This includes the 2-value for the direct estimate. The standard
deviation for the distance between events 4 and 7 amounts t o 0.153 (see
Table 6.4). This is nearly equal t o the value of the interval itself ( = 0.157).
It would indicate that the latter is not significantly different from zero. A
rapid test of this hypothesis (approximate t-test) consists of multiplying
the standard deviation by 2 and subtracting the result from the estimated
distance. If the difference is negative, the distance could well be zero.
Application of this test to the values listed in Table 6.4 shows that only 3 of
the intervals computed for the Hay example would be greater than zero
with probability greater than 95 percent. Equation (6.13) is based on the
assumption t h a t the xi-values a r e realizations of stochastically
independent random variables. This condition may not be satisfied in
practice and the estimated standard deviations may be too small.

When all possible comparisons can be made as for the pair of events 4
and 7, N* = N-1 where N denotes total number of events. However, in the
RASC computer program, N* may be less than N-1 for the following two
reasons: (1)The total number of comparisons is reduced by one for each
value xi that cannot be computed because one of the 2-values needed is
missing (this includes the case that both 2-values are missing); (2) if
Sij = Rij, Pij = 1 and the corresponding 2-value is set equal t o the
threshold value qc ( = 1.645 in Table 6.2). Pairs of 2-values both equal to
q,, and with zero-difference, are not used for estimating the average
distance A,q unless a pair of this type is contained within a cluster of
mutually inconsistent events. For this reason, pairs of values (Zjk, Zjk) in
successive columns (i, j = i + 1) are tested by letting h decrease from
h = i+ 1. Suppose that, for a given value of h , 2 i k = 2 j k = q,. This pair is
not used for the distance estimation unless a pair of 2-values, which are
not both equal to q,, is found for a smaller value of h . In the RASC
program, it is assumed that this situation is encountered as soon as five
pairs of 2-values equal to q, have been identified for decreasing h .
198

Likewise, pairs of values ( Z i k , Z j k ) in successive rows can be tested by

+
letting k increase from k = i 2.
Both preceding situations occur in the Hay example for estimation of
the distance between events 8 and 6. Because the 2-values for these
events combined with event 9 both are equal to qc = 1.645 (see first row of
Table 6.2B), and because the pair (8, 6) also has two non-determined
values, N* = 9 - 3 = 6. The corresponding weight (W) in Table 6.4 is only
4.7. The standard deviation ( = 0.203) for the corresponding interval
( = 0.770) is relatively large. Nevertheless, application of the preceding
approximate t-test suggests t h a t the latter value is statistically
significant.

When a large number of events for a long time interval is used, N* is

likely t o be much smaller than N-1 in all distance calculations, because
events belonging to relatively young assemblages (e.g. Late Miocene in
Fig. 6.2) normally all occur above events in older assemblages (e.g. Early
Eocene in Fig. 6.2). Distance estimates based on few pairs of 2-values are
relatively imprecise. In the RASC program there is a n option t h a t
distances based on N* less than m,2 are replaced by zeros.
The choice of a value for qc usually is not critical, because most pairs
of q,-values will not be used for distance estimation. D’Iorio (1990) has
performed a study of the effect of systematically changing qc for his data-
base (cf. Section 8.2). The average distance between successive events
increases when qc becomes larger but, in general, the relative order of the
events is not changed significantly. As a “default”, qc is set equal to 1.645
in the RASC program. This corresponds t o a cross-over frequency of
P = 0.95 (see Table 6.1). The user can replace the default value by any
other value. In general, qc should be greater than 1 and less than 2. It
should be kept in mind that the value of qc is selected because,
theoretically, a cross-over frequency of 1 corresponds to an infinitely large
2-value and distance estimation would not be possible. It can be assumed
that the scores from which cross-over frequencies are calculated satisfy
binomial frequency distributions. For small samples, the probability that
a cross-over frequency is equal to 1 (or 01, then is relatively large even
when a minimum sample size (m,p) has been defined. This problem is
restricted t o the tails of the normal (Gaussian) frequency curve and can be
solved by choosing a q-value which, effectively, changes the range of the
normal curve from (- -, -) to (-qc, q,).
 199

Reordering of events in the scaled optimum sequence

The last interval estimated in Table 6.4 is negative. For this reason,
it is desirable to reorder the events before a dendrogram of successive
interfossil distances is constructed. The cumulative distance from the first
event (No. 9) in the original optimum sequence obtained by ranking can be
calculated for each event in weighted as well as unweighted distance
analysis. In Table 6.4, the distance between events 9 and 2 (2.149) is less
than that between 9 and 3 (2.155). If distances from event 9 are used, it
follows that event 2 should lie above 3 in the scaled optimum sequence.
The events always can be reordered on the basis of this cumulative
distance. This allows the clustering of successive distances as shown, for
example, in Figure 6.2.

The standard deviations of the distances between successive events

cannot be recalculated readily after a reordering which removes negative
distances. This is because successive distance estimates a r e not
stochastically independent. In order t o obtain the new standard
deviations, it is necessary t o repeat all calculations taking the reordered
optimum sequence as the starting point. Because different Z-values then
are used for estimation, the distance estimates will change as is illustrated
in Table6.5 for the Hay example. New negative distances may be
computed a t this stage and the procedure would have to be repeated again.
These new calculations can be performed by using the final reordering
option of the RASC program. The objective of final reordering is to obtain
a set of distances between successive events which are all positive so that
the corresponding standard deviatons also are known. This result readily
could be achieved for the Hay example. However, when the data base is
large, and when h, and m,2 are small, it may not be possible t o obtain a
single set of consecutive distances which are all positive. This is because
the iterative process does not necessarily converge to a single solution. As
a default, at most four complete reorderings are allowed in the RASC
program. If convergence to a situation of positive distances is not obtained
in four or more steps, either the result without final reordering can be
accepted, or the result obtained after four or more reorderings. In the
latter solutions, the number of negative distances probably will have been
reduced considerably.

Figure 6.7 illustrates that the preceding iterative process for final
reordering does not necessarily converge to a single solution. Suppose that
the numbers in Figure 6.7 represent estimated distances between pairs of
200

TABLE6.5

Example of weighted distance analysis after reordering. The optimum sequence used as input for scaling
was not the ranking result used for Tables 6.2 to 6.4 but the scaled optimum sequence in the ranking of
events in last column ofTable 6.4. Differences between Tables 6.4 and 6.5 are restricted to values in two
rows at the bottom only.

Events Interval ~(x) Distance

1 9-10 0 317 0 100 0 317
2 10-8 0 176 0 289 0 493
3 8-6 0 770 0 203 1263
4 6-4 0 266 0 163 1530
5 4-7 0 157 0 153 1686
6 7-5 0 157 0 085 1843
7 5-1 0 195 0 082 2 038
8 1-2 0 118 0 147 2 156
9 2-3 0 006 0 124 2 162

B n

A @c ;@
A

4
I 2
0 4 3 2

E 3 D E 2 c

ABCDE -< ADBCE

Fig. 6.7 Artificial example for demonstrating that the final reordering option of the RASC computer
program does not necessarily converge to unique solution. See text for further explanation.

events A, B, C, D and E. A positive distance from one event t o another is

indicated by an arrow pointing from the one event t o the other. For
+
example, the distance from A to B is 2 and that from C to D is -2. Let the
optimum sequence ABCDE have only one negative distance (between C
and D). Because this distance is greater than that between B and C
( = + l), the reordered sequence becomes ADBCE. The distances for this
 201

artificial example have been chosen in such a way that this new sequence
again has only one negative distance (between D and B) and reordering
ADBCE gives the original sequence ABCDE. Consequently, a unique
solution with positive distances between successive events does not exist.
Situations similar t o the one illustrated in Figure 6.7 do occur in practice,
especially in situations where the estimated distances are not very precise.

6.4 Artificial example

The purpose of this section is t o illustrate the theory of scaling as

developed in the previous section by using the artificial example of Table
4.12 based on random normal numbers. Although the theory leads to valid
results for large samples, small-sample fluctuations may be considerable.
This aspect will be evaluated here. In general, the understanding of
statistical models applied to observed data can be helped considerably by
simulation experiments. Nevertheless, it should be kept in mind that
numbers are used, of which it may be known beforehand that they should
fit well because all expected values were determined by the scientist
conducting the experiment. In practical applications to real data, the
conditions artificially created for a simulation experiment may not be
satisfied. The artificial example of Table 4.12 clearly demonstrates some
features of the theory outlined in the previous section. However, it differs
from natural situations by (1)small number of events, (2) large number of
sequences, (3) all events are observed in all sequences, and (4) the
positions of the events satisfy normal distributions with equal variances.

By counting, it was determined that A is followed by B in 116 of the

150sequences of Table 4.12. (This implies t h a t B precedes A i n
34 sequences.) Likewise, A precedes C in 130 sequences and B comes
before C in 85 sequences. These three numbers ( n )are shown in the first
column of Table 6.6. They were transformed into relative frequencies (f,
by dividing them by 150 (see column 2 of Table 6.6). By consulting a table
of cumulative frequencies for the normal distribution in standard form,
the f-values were converted into 2-values. Multiplication by d 2 then
yields direct estimates of the distances between the events. For example,
DAB = 0.750 d 2 = 1.061. Only one indirect estimate of the distance DAB
can be obtained. It is equal to 1.335 which represents the difference
between 1.571 (direct estimate of DAC) and 0.236 (direct estimate O f DBC).
The arithmetic average of the direct and indirect estimates of distance is
202

TABLE6.6

RASC method of scaling applied to data of artificial example. For meaning of column headings, see text.

I1 I I I) (dircci) 11 (indirect) l h (Awl 1) (Ave) F (D)

AB 116 0.7733 0.7w mi 1.335 1.198 1.152 I .ooo

Ar 130 0.8667 1.111 1.571 1.297 I.d74 I,480 1.500

HC xs 0.5607 0.167 0.736 0.510 0.373 0.327 0.m

SSD 0.079 n.152 n.w 0.053

~~

shown in column 6 of Table 6.6. This is followed by a weighted distance

+
estimate which satisfies D(Ave) = (1.061 1.335/2)/1.5 = 1.152. Finally,
the expected value E D ) is shown in the last column.

Comparison of the three estimates of distance is facilitated by

computing the sum of squared deviations (SSD)from the expected value of
each estimate. For example, for the direct estimate of D in Table 6.6,

SSD = (1.061 - 1.000)2 + (1.571 - 1.500)2+ (0.236 - 0.500)2 = 0.079.

The SSD values also are shown in Table 6.6. The results suggest that
the variance of the indirect estimate which is proportional t o its SSD value
is about twice as large as that of the direct estimate. The weighted
average distances are most precise because they have the smallest
variance.

The analysis shown in Table 6.6 was repeated for the 5 smaller
subsamples. In all instances, the weighted mean distance provided the
best estimate (see Table 6.7). It also can be seen, however, that in small
samples, the estimated distance may differ considerably from its expected
value.

In the preceding statistical analysis of which the results are shown in

Tables 6.6 and 6.7, the weighted distance estimate D(Ave) was obtained by
assigning twice as much weight t o the direct estimate as to the indirect
estimate. Because weights are inversely proportional to variances, this
simply reflects the fact that, on the average, the variance of indirect
estimates is twice as large as that of direct estimates (see Fig. 6.6).

Suppose, however, that the equation for estimating weighted

distances Aei, i+ 1 as in the RASC program is used. From the values in the
 2 03

TABLE6.7

Statistical analysis of Table 6.6 repeated for the five subsamples.

Su bsa rnple D 1) Do D
(direct) (indirect) (Ave) (Ave)

I ... AH 1.030 1.089 1.060 1.050

AC 1.571 1.512 1.532 1.551
I%<' 0.482 0.541 0.512 0.502
SSD 0.006 0.010 0.006 0.006
2 ... AR 0.741 0.911 0.826 0.798
AC 1.030 0.860 0.945 0.973
TIC' 0.119 0.289 0.289 0.176
SSD 0.459 0.462 0.426 0.424
3 ... AB 1.191 1.809 1.500 1.112
AC 1.571 0.953 1.262 1.365

BC -0.238 0.380 0.071 0.032

SSD 0.580 0.968 0.491 0.314
4 ... AB 0.X81 2.236 1.559 1.333
AC 2.594 1.239 1.917 2.142
BC 0.358 1.713 1.036 0.810
SSD 1.231 3.067 0.774 0.619

5 ... AB 1.571 1.089 1.330 1.410

AC 1.571 2.053 1.812 1.732

BC 0.182 0.000 0.241 0.321

SSD 0.331 0.564 0.273 0.254

columns for fand 2 in Table 6.6, it readily is computed that WAB = 77.593,
WAC = 60.139, and WBC = 94.549. On the other hand, WAB.C = 36.758,
WAC.B = 42.618, and WJJC.A = 33.880. The latter three weights are not
exactly half as large as the first three weights. The reason for this
discrepancy is that the values of f a n d 2 are approximations only. They
were estimated from samples and used instead of the population values in
the RASC method. The RASC weighted distances become 1.149,1.457 and
0.308, instead of 1.152, 1.480 and 0.327 shown in Table6.6 for D(Ave).
Their correponding SSD value becomes 0.061 indicating that the D(Ave)
2 04

values (with SSD = 0.053) of Table 6.6 are better in this artificial
example.

6.5 Computer simulation experiments

As already pointed out in Section 4.9, the type of experiment described

in the previous section can be performed on a computer using a pseudo-
random number generator. Computer simulation experiments on
biostratigraphic events also have been performed by Edwards (1982) and
Harper (1984). Edwards (1982) dealt with the problem previously
illustrated in Figure 2.19. She assumed that a taxon has a true extinction
point in time which was randomly displaced upwards or downwards in a
section due t o sediment mixing. The aim of experiments of this type is to
model the distribution of exits (or entries).

Harper (1984) performed computer simulation t o create artificial

successions of taxa in sections. Three types of optimum sequences were
obtained for exits and entries of these taxa by means of the RASC
program: (a)presorting option only; (b) modified Hay method only; and (c)
scaling sequence as derived from (b). Harper demonstrated that, for his
successions, (a) consistently gave results that were better than those of (b)
and recommended that(a) instead of(b) be used as the input sequence
for (c). However, results of (b) could be improved by making m,l as small
as possible (see Section 7.4). A revision made in the RASC program on the
basis of Harper’s results was to allow the usage of different threshold
values for number of pairs of events (m,l and m,2; also see before). These
examples illustrate that computer simulation experiments can provide
results which are useful because they are complementary t o results
obtained from the analysis of natural data sets.

The computer simulation experiments used to test RASC were as

follows. Twenty artificial stratigraphic events were generated for each of
50 sections whereby the interval between the expected positions of the
events was kept constant in each run. For comparison, in the experiment
of Table 4.12, three events were studied in each of 150 sections and, in
total, 3 x 150 = 450 random normal numbers were used. In addition,
twenty artificial stratigraphic events were generated for each of 25
sections using only 3 spacings as described in Section 4.9. The latter
abbreviated database was shown previously in Tables 4.13 t o 4.15. For
each full experiment to be described here, 20 X 50 = 1000 random normal
 205

numbers with u = 1 were used. Every experiment was performed twice

using a different set of 1000 random normal numbers but the same set of
1000 random normal numbers was used in each of two sets of experiments
where the interval between expected positions E(D) was set equal to 1.0,
0.5,0.3,0.2,0.1 and 0.0, respectively. For scaling the threshold parameter
g , was set equal to 2.326 corresponding to P=O.99 which is midway
between 49/50 and 50/50. The latter two values are the largest frequencies
P which are possible in the database for 50 sections (without ties). For the
abbreviated database for 25 sections, q c was set equal t o 2.054
corresponding to P= 0.98.
Table 6.8 shows the artificial sequences of events created for the first
two sections for all 6intervals in one of the two sets of experiments.
Table 6.8 illustrates that for E D ) = 1.0 (u = 1.01, there are relatively few
inconsistencies in the observed sequences. On the other hand, for
E(D) = 0.1 it is difficult to recognize from the sequences that the expected
sequence is 1, 2, ..., 20. The sequences for E(D) = 0.0 are, of course,
completely random.
Table 6.9 shows optimum sequences obtained by (a) presorting option;
(b) ditto, followed by modified Hay method; (c) scaling (unweighted
differences); (d) scaling (weighted differences); and (e) ditto, after final
reordering.

TABLE 6.8

First two artificial sequences used in complete set of computer simulation experiments (20 events in 50
sections) with E(D) equal to 1.0,0.5,0.3,0.2,0.1,and 0.0, respectively.

Series Expected Sequence Numbers

No. C(n) I 2 3 4 5 6 7 8 9 10 II I 2 13 14 I5 16 17 18 I9 20

I 1.0 I 2 4 5 3 6 8 10 9 7 II I2 13 14 15 17 16 18 19 20

2 1.0 I 4 3 2 7 6 5 8 9 I1 10 12 I3 15 14 16 18 17 19 20

I 0.5 I 2 5 4 3 6 10 8 9 I1 I3 I4 12 15 7 17 16 18 19 20
2 0.5 I 4 3 2 7 8 9 6 I1 5 I2 13 10 I5 18 19 16 14 17 20

I 0.3 5 I 4 2 10 3 6 8 I1 9 15 I3 14 17 12 16 7 19 18 20
2 0.3 1 4 3 7 2 8 9 II 6 I2 13 18 I5 5 10 19 16 20 17 14

I 0.2 5 4 I 2 10 II 3 6 8 9 15 17 14 13 12 16 19 18 20 7

2 0.2 I 4 7 3 I1 8 9 13 18 I2 2 I5 6 19 20 10 16 5 17 I4

I 0.1 5 10 4 2 I II 17 I5 I4 8 9 13 6 3 16 I2 19 20 I8 7

2 0.1 I 4 7 18 II 19 9 13 8 12 3 I5 20 2 6 16 17 10 14 5

I 0.0 5 10 17 IS II 4 14 13 I6 19 2 9 20 8 I I8 6 I2 3 7

2 0.0 I I8 19 4 20 II 13 15 7 I2 9 8 17 16 3 10 6 14 2 5
206

TABLE6.9

Optimum sequences obtained by 5 methods (a to e ) of ranking applied to 5 artificial sequences (I to V) of

events, Mean intervals were 1.0 (I), 0.5 (II), 0.3 (III), 0.2 (IV) and 0.1 (V), respectively. A and B denote
results obtained for two different datasets (1 and 2).

From the results in Table6.9, it can be concluded that all optimum

sequences are either equal t o or close to the expected sequence. However,
it cannot be decided which optimum sequence is best (also see Section 7.2).
Results obtained from these computers simulation experiments will be
further discussed in Chapter 8. In the remainder of this section, the
smaller database originally presented in Section 4.9 will be used. It is
sufficiently large to show that (1)the scaling method presented earlier in
this chapter provides unbiased estimates of intervals between successive
events, and (2) the standard deviations computed by means of Equation
(6.13) may be too small because the direct and indirect distance estimates
that are being averaged are not stochastically independent.

Scaling of abbreviated database.

The original SEQ files for this data base were shown in Tables 4.13 to
4.15. All three sets were subjected to probabilistic ranking (presorting)
followed by modified Hay method. The resulting optimum sequences and
dendrograms (obtained after final reordering) are shown in Figures 6.8 to
6.10 (from Gradstein et al., 1985). As pointed out before, an identical
 207
1 0 1915
3 0 0153
SEQUENCE fOSSIL
POSITION NUMBER 2 0 20011
1 0 - 2 I 0 4.817
2
3
4
5
1 -

2 -
3 -
4 -
3
4

5
6
1 5
6
7
8
0 2032
0 4502
0 2263
0 4360
6 5 - 1 9 0 3653
1 6 - 8 10 0 1063
8 7 - 9 11 0 3099
9 I - 10 12 o 4ng3

10 10 9 - 11 13 0 0571
11 10- 12 14 0 4810
12 11 - 13 15 0 0266
13 13 12- 14 16 0 8437
17 n 0487

,
14 I4 13- I5
15 15 1,- 18 1.8 n 5037
16

17
16
17
15-
1.8-
17
18
I 19
20
0 6017

I8 1.8 17 - 19
19 19 1.8- 20
20 20 19- 21
INTERFOSSIL DISTANCES

Fig. 6.8 Optimum sequence and dendrogram for sample drawn at random from population (theoretical
model of equally spaced events labelled 1 to 20 along RASC scaled with E(D)=0.5). Original SEQ file is
shown in Table 4.13.

sequence of random normal numbers was used in all three experiments.

Consequently, there is a general resemblance of input and output for the
three SEQ files. The deviation from the theoretical model, consisting of
equally spaced events ordered 1 to 20, increases when E(D), representing
spacing between successive events in the theoretical model, decreases.

The modified Hay method did not change the probabilistic ranking
result in the experiments with E(D)= 0.5 or 0.3. Probabilistic ranking
results were changed somewhat by subsequent application of the modified
Hay method for E(D)= 0.1. Eight 3-event cycles occurred and were broken
by temporarily zeroing the (first) pair of elements in the cycle with the
smallest difference as shown in Table 6.10. In total, seven (12, 13) pairs
and one (11,14)pair were ignored in order t o obtain the optimum sequence
of Figure 6.10. However, a detailed comparison of the probabilistic
ranking result (Table 6.11A) with the final optimum sequence (Table
6.11B) shows that the probabilistic ranking is closer to the true sequence
than the modified Hay ranking. The numbers in the bottom rows of Tables
6.11A and B are absolute values of differences between ranking results
and true order numbers. Their sum is 22 for Table 6.11A and 33 for Table
6.1 1B suggesting that the probabilistic ranking result is slightly better in
this type of application (also see Section 7.4). For further comparison,
Table 6.11C shows differences for the scaled optimum sequence of Figure
 1 0 1391
3 0 1123
SEWENCE FOSSIL

'
RANCE 2 0 0887
PO51 TI ON NUMBER
4 0 2411
1 1 0 - 2
3 1 - 3 5 0 0792
2
6 0 2671
3 4 2 - 4
7 0 2540
4 2 3 - 5
5 5 4 - 6 8 0 2801
9 0 3117
6 6 5 - 7
7 7 6 - 0 1 1 0 0291

8 8 7 - 9 10 0 1780

9 9 8 - 10 12 0 1800

10 10 9 - 1 1 14 0 0998
13 0 2940
1 1 11 10- 12
12 12 1 1 - 13 16 0 0050

13 14 12- 14 15 0 4702

14 13 13- 15 17 0 0783

15 16 14- 16 18 0 3302

18 15 15- 17 19 0 4130
I 1 20
17 17 16- 18
18 18 17- 10
19 19 10- 2 0
20 20 1 s - 21
INTERFOSSIL DISTANCES

Fig. 6.9 Optimum sequence and dendrogram for computer simulation experiment of Fig. 6.8 repeated
with E(D)=0.3. Original SEQ file is shown in Table 4.14.

4 0 1171
3 0 0546
SEQUENCE FOSSIL
RANCE
POSITION NUMBER 1 0 0131
1 4 0 - 2 5 0 0190
2 5 1 - 3 6 0 1015
3 3 2 - 4 2 0 0693
4 1 3 - 5 7 0 0783
5 2 4 - 6 8 0 1806
6 6 5 - 7 9 0 0293
7 8 6 - 8 1 1 0 0621
8 7 7 - 9 16 0 0104
9 1 1 8 - 10 14 0 0664
10 9 9 - 1 1 10 0 0483
1 1 14 lo- 12 12 0 0121
12 16 1 1 - 13 15 0 0632
13 10 12- 14 13 0 1676
14 12 13- 15 17 0 0155
15 15 14- 16 18 0 0953
16 13 15- 17 19 0 3337
I
17 17 16- 18
18 18 17- 19
19 19 18 - 2 0
20 20 19- 21
INTERFOSSIL DISTANCES

Fig. 6.10 Optimum sequence and dendrogram for computer simulation experiment of Fig. 6.8 repeated
with E(D)= 0.1. Original SEQ file is shown in Table 4.15.

6.10. These add t o 28 which is about midway between the preceding two
sums.
 209
TABLE 6.10

Eight 3-event cycles detected during application of modified Hay method (after probabilistic ranking) to
SEQ file shown in Table 4.15.

r
A B c D

4 3 5 2 1 6 7 6 2 10 9 11

x 11 14 x 11 14 x 11 13 x 9 13
14 x 10 14 x 12 14 x 11 16 x 9
11 15 x 11 13 x 12 14 x 12 16 x

E F G H

14 9 11 16 14 12 10 16 14 16 13 12

x 11 13 x 10 16 x 10 13 x 12 16
14 x 9 15 x 12 1s x 10 13 x 12
12 16 x 9 13 x 12 15 x 9 13 x

TABLE 6.11

Comparison of true optimum sequence with optimum sequences resulting from (A) probabilistic ranking,
(B) modified Hay method after probabilistic ranking, and (C) scaling after probabilistic ranking and
modified Hay method. Absolute values of differences between estimated and true ranks can be regarded
a s penalty points. In the RASC step model (Chapter 7) these penalty points will be added to obtain a
statistic from which Kendall's rank correlation can be computed.

Estimates of intervals between successive events in the optimum

sequence using both unweighted ($1) and weighted (22) scaling analysis
are shown in Tables 6.12 to 6.14 for the three experiments, respectively. It
can be seen t h a t N* tends to decrease towards the top and the base of the
composite section if E(D) increases. This is because of reduced probability
of inconsistencies between events near top and base, respectively, which
leads t o pairs of 9,-values not used for distance estimation (see before).
210

Standard deviations obtained by Equation (6.13) for the weighted

scaling results are shown in the last columns of these three tables.
Because the standard deviation of the normal random numbers used in the
computer simulation experiments was equal to unity (instead of equal to
(2)-* as in RASC), E(D)must be divided by d 2 in order to obtain expected
values for 32. For example, the expected value for all numbers in the
fourth (XI) and fifth (52) column of Table 6.12 is 0.5/d2=0.354. The
corresponding expected values for Tables 6.13 and 6.14 are 0.212 and
0.071, respectively. However, in these two tables the order of the twenty
events is not 1 to 20 as in Table 6.12. For example, in Table 6.13 event 1 is
follwed by event 3 instead of event 2. Consequently, the expected value for
this interval is 2X0.212 instead of 0.212. Mean intervals computed from
19 separate distances between successive events are shown at the bottom
of each table. In all instances, the mean is close to the expected value of
the interval between events with consecutive numbers, although the
scatter of the individual 31 and X2 values is considerable. When the
TABLE 6.12

Unweighted (21) and weighted ( 2 2 ) estimates of intervals in scaled optimum sequences for computer
simulation experiment with E(D)=0.5. Standard deviations s ( i 2 ) of weighted estimates are shown in
last column.

Events N* LI X? s! i.L)

1 1~2 9 n 455 n 427 0 083

2 2-3 9 0 038 n 015 o 093

3 3-4 7 0 192 0 242 0 06fi
4 .1-5 in 0 541 0 482 o 079
5 56 I0 0 181 0 203 0 081
6 6-7 14 0.466 0 150 n 0133
7 7-8 13. 0 169 0 22fi 0 071
8 8-9 14 0.391 0 136 n 081
9 9-10 14 0.288 0 34fi !i082

in 10~11 14 an96 0 1111 0 034

11 I I 12 13 0 297 !i310 0 075

12 12-13 11 0 907 0 409 0 051

13 13-14 13 0.049 n 057 0 047

14 14-15 13 0.420 0 181 0 074

15 1516 13 o 030 0 027 0 039

16 16 17 11 0 7fifi 0 844 0 115

17 IT18 8 0 083 n 099 0 161

18 18-19 8 0 470 0 501 0 065

19 19-20 fi 0 53fi 0 (in8 n 113

 21 1

TABLE 6.13

Same as Table 6.12 for E(D)= 0.3

I 1-3 12 0 ll0 0 139 0 072

2 3-4 15 0 207 0 209 0 062

3 4-2 15 0 122 0 089 0 078

4 2-5 17 0 335 0 346 0 056

5 5-6 17 0 012 0 079 0 075

6 6-7 16 0 274 0 267 0 046

7 7-8 17 0 281 0 254 0 064

8 8~9 17 0 259 0 280 0 068

9 9 10 19 0 329 0 338 0 ofi2

10 10-11 I9 -0 042 0 029 0 068

II 11-12 19 0 194 0 222 0 077

12 12-14 18 n 183 0 180 0 059

13 14-13 18 0 061 0 100 0 065

14 13-16 17 0 308 0 283 0 068

15 16-15 17 0 053 0 005 0 063

16 15-17 17 0 444 0 470 0 059

17 17 18 15 0 017 0 078 0 066

I8 18 19 15 0 320 0 330 0 073

19 19~211 I0 0 309 0411 0 088

reordering of events in Tables 6.13 and 6.14 is considered, the mean

expected values for these two tables are increased t o 0.302 (instead of
0.212) and 0.112 (instead of 0.0711, respectively.

Table 6.15 contains summary statistics for the three data sets.
Separate standard deviations (s and 6) were computed from the samples of
19 values with respect t o the sample mean (unbiased estimate, 18 degrees
of freedom) and the population mean (unbiased estimate, 19 degrees of
freedom). For E(D)= 0.5, the pairs of estimates (s and 6 ) are nearly equal
t o one another. For E(D)= 0.3 and 0.1, &(XI) and 8322) are larger than ~(321)
and ~ ( 3 2 2 because
) the order of the events in the corresponding optimum
sequences differs from the expected order (1, 2, ..., 20). Ordinary product-
moment correlation coefficients r ( f 1 ,i 2 ) are shown at the bottom of Table
6.15. They indicate that the unweighted and weighted analysis results are
strongly correlated.
212

TABLE 6.14

SameasTable6.12forE(D)=0.1

1 4-5 19 0.207 0 182 0 063

2 5-3 19 -0 103 0 064 0 066

3 3- I 19 0.035 0 055 0 060

4 1~2 19 0.146 0 133 0 044

5 2-6 19 ~0.126 0 102 0 055

6 6~8 19 0 266 0 250 0 053

7 8-7 19 -0.081 0 078 0 040

8 7 I1 19 0.361 0 286 0 067

9 11-9 19 ~ In) f i i 0 029 0 064

10 9 14 I9 0.100 0 103 0 063

11 14-16 19 -0.019 0011 0 059

12 16-10 19 0 064 0 077 0 062

13 10.12 19 0 038 0 048 0 048

14 12-15 19 0 058 0 072 0 039

15 15-13 19 0 055 0 063 0 053

16 13~17 19 0 19fi 0 168 0 071

17 17~18 19 -0,008 0 016 0 048

18 18~19 19 0.117 0 095 0 055

19 19-20 19 0 285 0 334 0 069

Mean Interval. 0.078 0 084

The standard deviations s(X2) and 6(22) as computed from the 22

values are considerably greater than individual s(i2) values listed in the
last column of each table. The explanation of this discrepancy is that the
N* separate xg-values, on which f 2 and s(X2) are based, are not
stochastically independent. (Two random variables are stochastically
independent if the expected value of the cross-product for deviations from
their means is equal to zero. The expected value of the product-moment
correlation coefficient of two stochastically independent variables also is
zero). The computer simulation experiments show that the effect of
mutual interdependence obviously can be strong for smaller as well as for
larger expected intervals between successive events. It does not lead to
noticeable bias in the %values but may result in standard deviations
which are 2 or 3 times too small.
 213

TABLE 6.15

Comparison of estimates of Tables 6.12 to 6.14 to population parameters. See text for explanations of
expressions in first column.

0.354 0.212 0.071

0.296 0.180 0.078
0.326 0.204 0.084
0.227 0.160 0.132
0.228 0.194 0.193
0.225 0.152 0.121
0.221 0.203 0.194
0.981 0.980 0.987

Mutual interdependence of xi
The following considerations a r e helpful for understanding t h e
mutual interdependence of separate estimates xi of the interval between
two events A and B. Let C and D be two other events which can be used for
indirect estimation of the distance between A and B.
Three events A, B and C are related by the six probabilities PABC,
PACB,PCAB,PBAC,PBCA and PCBA. It follows immediately that

‘BC = ‘ABC + ‘BAC + ‘BCA (6.14)

Similar expressions can be written out for PBA, PCA and PBC but it is
simpler to regard these probabilities as complementary to PAB, PAC and
PBC, or
P, = 1-PAB; PCA = 1 -PAC; PCB = 1 -PBc
(6.15)
2 14

In Section 6.1 it was pointed out that the 2-value of a probability P

can be approximated by a 2*-value which simply is a linear function of P,
or 2*=2.93(P-0.5). This relation can be used t o prove that the indirect
estimate ZAB.C=ZAC-ZBCis approximately equal to

'*AB.C - '*AcB - '*BCA (6.16)

where Z*ACB and Z*BCA are linear functions of PACB and PBCA,
respectively. Consequently, for the two events C and D, it follows that

ZABC 2.93 (PACB-PBCA)

(6.17)

The latter two values cannot be regarded a s stochastically

independent estimates of the interval between A and B. Even if C and D
are independent of one another, as well as from A and B, the estimates of
Equation (6.17) both depend on the positions occupied by A and B in the
sections.
In the computer simulation experiments, all events are present in all
sections. The preceding argument explains why Equation (6.13) provides
poor results in that situation. In most practical applications of ranking
and scaling, the events do not all occur in all sections. As pointed out
before (cf. Fig. l.l),frequency distributions of events usually are positively
skew with most events occurring in relatively few sections. If the two
events C and D in the preceding argument occur in different sections, the
two estimates of Equation (6.17)would be stochastically independent.

Equation (6.13) only would provide a n approximately unbiased

estimate in the unlikely situation that all events used occur in different
sections. In practice, it can be expected that the estimated standard
deviations will be too small, although bias will not be as severe as in the
computer simulation experiments. This type of bias also must be
considered in results of the normality test (see next section). In that
situation, the computer simulation experiments initially helped to point
out the problem. This led to a revised normality test in which the bias
could be estimated and eliminated i n practical applications (see
Chapter 8).
 215

6.6 Normality test

When a n optimum sequence has been obtained, individual sequences
for sections can be compared to it. In practice, this is important as a tool t o
spot anomalous events which occur either much higher or much lower
than expected in a section. The RASC computer program provides three
optional outputs for ranking or scaling results immediately preceding the
normality test which applies t o scaled optimum sequences only. These are:

(1) An occurrence table can be constructed with the final ranking plotted
in the vertical direction. Sections in this table are represented by
columns. If an event occurs in a section its presence is indicated by
an X.
(2) Each section can be compared to the optimum sequence by using a
system of scoring penalty points when an event is out of place. This
procedure is called the Step Model (cf. Section 7.3). The relative order
of every pair of events is checked against their order in the optimum
sequence. If the order is different, one penalty point is scored. Coeval
events each receive half a point. Obviously, an event with many
penalty points is likely to be either too high or too low in a section. A
drawback of the step model is that events which belong to clusters of
events with many internal inconsistencies are likely to accumulate
high total scores even if they occur in normal positions. Thus, it may
not be easy to distinguish between anomalous events which are out of
place and events which are part of a cluster.

(3) Scattergrams: the order of events in the sections is plotted against

the optimum sequence.
The preceding three outputs can be obtained before or after scaling.
The normality test can only be performed after scaling. Two outputs for
the normality test are shown in Tables 6.16 and 6.17 for the Hay example.
After scaling, each event occurs at a fixed distance from the origin which
was set at the position of the first event in the scaled optimum sequence.
Hence, the score of each event, with the exception of the first and last
events, can be compared with the scores of its two neighbors in every
section. The amount by which it is out of place can be evaluated
statistically. For this purpose, the second-order difference will be used
216

TABLE 6.16

RASC normality test output for 9 sections of Hay example. E=event code number; L=event level
number in stratigraphically downward direction; X = cumulative RASC distance; and U = second order
difference. Single asterisk indicates that event is out of place with probability greater than 95%. Two
asterisks indicate that event is out of place with probability greater than 99%.

Section A - Vaca Valley E L X U

HI Discoaster tribrachiatus 9 1 0.000
LO Discoaster cruciformis 8 2 0.493 0.700
LO Discoaster minimus 7 3 1.686 -1.616
LO Rhabdosphaera scabrosa 6 4 1.263 1.683
LO Coccolithus gammation 5 4 1.843 -0.894
LO Coccolithus solitus 4 4 1.529 0.940
LO Discoaster germanicus 3 4 2.155 -0.632
LO Coccolithus cribellum 2 4 2.149 -0.105
LO Discoaster distinctus 1 4 2.038

Section B - Pacheto Syncline E L X U

HI Discoaster tribrachiatus 9 1 0.000
LO Discoasterdistinctus 10 2 0.317 2.778*
LO Rhabdosphaera scabrosa 6 2 1.263 -0.366
LO Caccolithus gammation 5 2 1843 -0.894
LO Coccolithus solitus 4 2 1.529 0.470
LO Discoaster minimus 7 2 1.686 0.313
LO Discoaster germanicus 3 2 2.155 -0.475
LO Coccolithus cribellum 2 2 2.149

Section C - Tres Pinos E L X U

HI Discoaster tribrachiatus 9 1 0.000
LO Discoaster distinctus 1 2 2.038 -2.234
LO Coccolithus gammation 5 3 1.843 0.502
LO Coccolithus cribellum 2 4 2.149

Section D - Upper Heliz Creek E L X U

LO Discolithus distinctus 10 1 0.317
HI Discoaster tribrachiatus 9 2 0.000 0.810
LO Discoaster cruciformis 8 3 0.493 0.857
LO Coccolithus gammation 5 4 1.843 -1.507
LO Discoaster minimus 7 5 1.686 0.509
LO Discoaster distinctus 1 6 2.038 -0.241
I,O Coccolithus cribellum 2 7 2.149

which is approximately equal t o the sum of the scores of the neighbors

minus twice the score of the event itself.

If the second-order difference value of an event in a section receives

two asterisks in the output, it is out of place with a probability of
99 percent. One asterisk signifies an event that occurs too high or too low
with a probability of 95 percent. It is t o be expected that, on the average,
one and five percent of all events tested will be assigned two asterisks and
 217

TABLE 6.16 (continued)

Section E -New Idria E L X U

HI Discoaster tribrachiatus 9 1 0.000
LO Rhabdosphaera scabrosa 6 2 1.263 -0.997
LO Coccolithus solitus 4 3 1.529 -1.303
LO Discoaster cruciformis 8 4 0.493 2.229
LO Discoaster minimus 7 5 1.686 -0.724
LO Discoaster germanicus 3 6 2.155 -0.586
LO Discoaster distinctus 1 7 2.038 -0.078
LO Coccolithus gammation 5 8 1.843 0.809
LO Coccolithus cribellum 2 8 2.149
Section F - Media Aqua Creek E L X (I
~~

LO Discolithus distinctus 10 1 0.317

HI Discoaster tribrac hiatus 9 2 0.000 0.810
LO Discoaster cruciformis a 3 0.493 1.044
LO Discoaster minimus 7 3 1.686 -1.074
LO Coccolithus cribellum 2 4 2.149 -0.770
LO Coccolithus gammation 5 5 1.843 0.337
LO Coccolithus solitus 4 5 1.529 0.595
LO Discoaster germanicus 3 6 2.155 -0.399
LO Discoaster distinctus 1 6 2.038
Section G - Upper Canada E I, X U
HI Discoaster tribrachiatus 9 1 0 000
LO Discoaster cruci formis 8 2 0 493 -0 248
LO Discolithus distinctus 10 2 0 317 1280
LO Coccolithus gammation 5 3 1843 -0 798
LO Coccolithus cribellum 2 3 2 149 -0 418
LO Discoaster distinct us 1 3 2 038 -0 819
LO Coccolithus solitus 4 4 1529 1556
LO Discoaster germanicus 3 4 2 155 -1 517
LO Discoaster minimus 7 5 1686
Section H - Las Cruces E L X U
LO Coccolithus solitus 4 1 1529
H I Discoaster tribrachiattrs 9 2 0 000 3 372**
LO Coccolithus gammation 5 3 1843 - 1 595
LO Discoaster distinctus 1 3 2 038 - 1 917
LO Discolithus distincfus 10 3 0 317 3 038*
LO Discoaster minimus 7 4 1 686
-
Section I - Lodo Gulch E L X (I

LO Discolithus distinctus 10 1 0 317

H I Discoaster tribrachiatirs 9 2 0 000 1580
LO Rhabdosphaera scabrosa 6 3 1263 -0 997
LO Coccolithus solitus 4 4 1529 0 047
LO Coccolithus gammation 5 5 1 843 -0 118
LO Discoaster distinct us 1 6 2 038 0 227
LO Discoaster germanicus 3 6 2 155 -0 428
LO Coccolithus cribellum 2 7 2 149
218

TABLE 6.17

Normality test applied to all ( = 51) second-order differences of Table 6.16. The expected frequencies (El
of the ten classes are all equal to 5.1. The chi-squared test is used for comparing observed (0) and
expected (E)frequencies with one another. Abnormally large values in the last column would be
indicated by asterisks (one and two asterisks for lack of fit with probability greater than 95% and 996,
respectively).
~

Class 0 h 0E (0E R E
1 3 51 21 0 39
2 5 51 ni 0 00
3 8 51 29 0 74
4 7 51 19 0 32
5 5 51 01 0 00
6 3 51 21 0 39
7 5 51 01 0 00

8 7 51 19 0 12
9 3 51 21 0 39
10 5 51 01 0 00

Chi-squared = 2.56

one asterisk, respectively, when there are no anomalies as in the computer

simulation experiments described in the preceding section. The 51 second-
order differences listed in Table 6.16 for the Hay example show one double
asterisk (HI Discoaster tribrachiatus) in Section H and t w o single
asterisks. This is approximately as expected for a normal dataset of this
size and no clear anomalies are indicated.
The model which assumes that each event has a normal probability
curve with the same variance, can be used to estimate a set of expected
frequencies for all observed event locations. The second-order differences
then are also normally distributed. For convenience, this theoretical
distribution is divided into 10 classes with equal expected frequencies. In
total, 10 expected frequencies for 51 values are compared t o the
corresponding observed frequencies in the bottom part of Table 6.17 for the
Hay example.
In the last column of Table 6.17, the squared difference, divided by
expected value corrected for autocorrelation (see Chapter 8) is shown.
Each of these values is approximately distributed as chi-squared with a
single degree of freedom, and the total for 10 classes has seven degrees of
freedom. Probable chi-squared departures from normality are marked by
 219

asterisks. A more detailed explanation of the procedures followed in the

normality test will be given in Chapter 8.
It should be kept in mind that the normality test is based on the
assumption that all second-order differences have the same variance. This
condition is only approximately satisfied if individual events have
different variances. In the latter situation, the averaging process helps t o
stabilize the variance. It also is noted that the number of second-order
differences considered is two less than the number of events per section. It
is possible to consider slightly different types of second-order differences
by considering coeval events.

6.7 Marker horizon option of the RASC method

This section deals with an option of the RASC computer program in
which the location of chronostratigraphic (marker) horizons such a s
seismic markers or bentonite beds resulting from volcanic ash fall can be
considered. If events of this type can be correlated with certainty between
sections, they should be assigned zero variance along the relative time
scale used for estimating RASC distances, when they are considered in
conjunction with other stratigraphic events. In practice, this means that
marker horizons will be given more weight than other events in the
calculations. Marker horizons are entered like other stratigraphic events
in the SEQ file for the RASC program. However, they are identified as
special events in the input specifications for RASC.
The underlying statistical theory for marker horizons is explained in
Figure 6.11 which can be compared to Figure 6.5 earlier in this chapter.
The event A has been replaced by a marker horizon with zero variance in
Figure 6.11. This means that DABis normally distributed with mean AAB
and variance u2. The 2-value for relative cross-over frequency PAB has t o
be divided by d2 before it can be used as a direct estimate of the distance
between events A and B which is compatible with the direct estimate of
Figure 6.5 and the indirect estimates of Figure 6.6. Its weight also has to
be adjusted accordingly.
When indirect estimates such as AAB.C involving a marker horizon
are obtained, there are two possibilities: (1) the event used for the indirect
comparison (C) is a marker horizon; and (2) either A or B is a marker
horizon. In the first situation, DAC and DBC are normal with variance u 2
220

Fig. 6.11 Direct estimation of distance AAB between events A and B from relative cross-over frequency
when A is a marker horizon with zero variance. The variance of D A B is equal to the variance of event B.

(as DAB in Figure 6.11). Consequently, their difference can be assigned

variance equal to 2u2. The second case results in difference DAB.Cwith
variance30 2 .
The preceding theory of marker horizons is further illustrated by
using a modified version of the artificial example of Table 6.6. Table 6.18
shows results comparable to those of Table 6.6 under the assumption that
B is a marker horizon. As before, a table of sequences for A, B and C was
derived (cf. Table 4.12). However, this time all random normal numbers
for B were replaced by 2.000. For example, the first three event “distances”
of Table 4.11 previously were 1.422,0.130,2.732and, therefore, gave BAC.
This time they result in 1.422,2.000,2.732and, therefore, ABC. (For one
sequence, the locations of A and B became both equal to 2.000 and 0.5 was
added to the tallies of both AB and BA.)

TABLE 6.18

RASC method of scaling applied to data of artificial example. Event B is marker horizon. These results
should be compared to those shown in Table 6.6 where B, like A and C, had unit variance.

n f Z D(direct) D(indirect) Do(Ave) D(Ave) E(3)

AD 129.5 0.8633 1.095 1.095(1) 1.283(3) 1.189 1.142 1.000

AC 130 0.8667 1.111 1.571(2) 1.383(2) 1.477 1.477 1.500
BC 92 0.6133 0.288 0.288(1) 0.476(3) 0.382 0.335 0.500
SSD 0.059 0.094 0.050 0.048
 22 1

Comparison of Table 6.18 to Table 6.6 shows the following differences.

The cross-over frequencies for AB and BC have increased significantly.
This reflects the fact that the variance of B was set equal t o zero. The
direct estimates of distance for AB and BC were not multiplied by d 2 as in
Table 6.6. However, D (indirect) was estimated from D (direct) by simple
addition or subtraction as before. The relative variances of D (direct) and
D (indirect) are shown in brackets in Table 6.18. As before Do(Ave) is the
arithmetic average and D(Ave) represents the weighted average. The
weights are inversely proportional to the variances. For example, D(Ave)
+
of AB is equal to (1.095 1.283/3)/(1+ 1/3) = 1.142. The SSD values in
Table 6.18 are less than those in Table 6.6.

6.8 Unique event option of RASC program

The purpose of this option is that a rare (unique) event can be entered
into a regional standard by comparing its position in a single or few
sections t o those of the more abundant taxa used for constructing this
standard. The unique event option is useful when an index fossil is
observed in one or a few sections only. Because of its rarity, the index fossil
cannot be used for construction of the optimum sequence. However, it can
be fitted in with the other events afterwards and this may help t o define
assemblage (average interval) zones from the dendrogram. Like marker
horizons, the unique events have to be identified in the input specifications
of the RASC program.
The unique event option can be used t o solve t h e following
hypothetical problem. A feature is observed in two sections and it is of
interest t o determine whether this feature represents a single
stratigraphic event. The feature, then, can be entered by codes as a
different unique event for each section. The resulting positions of these
two unique events in the standard can be compared to each other and this
may be helpful for deciding whether the same event is present in both
sections.
Figure 6.12 shows the technique used for treating unique events. The
event A with distance "t, from the origin is observed immediately above
the unique event in a section. Event S (with xs) is coeval t o the unique
event and event B (with q,) occurs below it. Also shown on the left of
Figure6.12 are the locations X'b and X"b for two other events B' and B"
observed below B. As a first approximation, the unique event is assigned
222

- xs

-Xa

I
VI
.r
X
m

1 - z1+ +R

Fig. 6.12 Simple example to illustrate application of unique event option. A unique event was observed
in a single section simultaneous to the event S, stratigraphically below the event A and above the events
B, B' and B". The cumulative RASC distances of the latter five events are shown along the scale on the
left. The positions of S, A and B were averaged to obtain first approximation f l for the unique event. The
second approximation was based on the RASC distances of all events within the range R.

the position 21 representing the arithmetic average of xa, xs, and q-,. In
practical applications, S may be missing. (The special situation that A or
B is missing would occur only if the unique event were to occupy the first
or last position in a section.)
More t h a n a single e v e n t ma y be observed i n t h e positions
immediately above, simultaneous to, or below the unique event. Then, xa,
xs, or q, will be computed as averages for these events which, in turn, will
be averaged to estimate 21.
A range of 21 k (1/2)R can be defined for all events encountered
within the vicinity of 21 with a probability greater t h a n 5percent.
Because u2 = 0.5, (1/2)R = 1 . 9 6 ~= 1.386. The events i n the scaled
optimum sequence with locations i n the interval 21 k (1/2)R can be
identified. For the simplified example of Figure 6.12, these are the events
A, S, B, and B' (but not B"). For each event above the unique event in the
section (A), a value is computed which is the average of its location (x,)
+
and the value 21 (1/2)R. Similarly, for each event below the unique
 223

event (B or B'), a value is computed which is the average of its location (xb
or x'b) and the value 31 - (1/2)R. These average values which are shown as
arrows in the diagram on the right of Figure 6.12 are averaged together
with the values (x,) for events observed to be simultaneous with the
unique event. This gives the second approximation 32. If the unique event
occurs in more than one section, the preceding calculation is performed for
each section and the resulting values of 32 are averaged.
The choice of a range R in the method of Figure 6.12 is somewhat
arbitrary. However, the location of the second approximation 3 2 is
independent of R when the number of events within the interval
31 k (1/2)R remains constant. Although the unique event option
generally is used for the construction of biozonations from the scaled
optimum sequence, it also can be used in association with an optimum
sequence obtained by ranking. In that situation, the sequence numbers for
events in the optimum sequence are used as x-values and R is set equal to
a larger value (e.g. R =3.0).
Examples of using the unique event option t o include index fossils in
biozonations are given elsewhere (see e.g. Fig. 6.2). The following example
illustrates the concept of re-including an event that initially was excluded.
Event 6 in the Hay example occurs in 4 sections only. By setting the
threshold parameter h , equal t o 5 , it can be excluded from the
computations required for ranking and scaling. Table 6.19 shows
optimum sequences obtained from 8 events with later re-insertion of event
6. In both sequences, event 6 is positioned between events 8 and 4 (but
closer t o 4 than t o 8) as in the results previously obtained for the Hay
example.

6.9 Binomial and trinomial models for scaling

As already pointed out in Section 3.4,the RASC model for scaling can
be evaluated in terms of observed probabilities by which the events
succeed one another in the wells or were observed t o be coeval. The
relation between binomial and trinomial models will be considered in the
final sections of this chapter.
Suppose that two stratigraphic events (either entries or exits or one of
each) for different taxa are expressed as Ei and Ej. If Ei and Ej occurred
relatively close in geological time, it may be that Ei is observed t o occur
224
TABLE 6.19

Test of unique event option applied to Hay example. Event 6 which occurs in 4 sections only was
excluded by setting k , = 5 . Later it was re-inserted in the optimum sequence derived by ranking as well
a s in the scaled optimum sequence.

Event Ranking Scaling

9 0.00 0.000
10 1.00 0.317
8 2.00 0.493
6* 2.87 1.017
4 3.00 1.198
7 4.00 1.415
5 5.00 1.534
1 6.00 1.724
2 7.00 1.868
3 8.00 1.875

above Ej in some outcrop sections or wells, and Ej above Ei in others. It

also may be that Ei and Ej are locally coeval. In order t o avoid confusion,
events for pairwise comparison will be denoted by using the letter A
instead of E. Binomial models provide estimates of the probability of A1 or
that Ei occurs above Ej in a section. If this probability is written as PI, and
the probability of A2 (Ej occurs above Ei) as P2, then P2 = l-P1. As
originally pointed out by Edwards and Beaver (19781, only a trinomial
model can result in an estimate of the probability of occurrence of A3 or
that Ei and Ej are observed to be coeval in a section in addition t o the
probabilities that A1 and A2 occur. The development of trinomial models
is of importance because biostratigraphic events are frequently observed
to be coeval and this possibility should be considered in the statistical
models.
Statistical theory of the binomial and trinomial distributions can be
found in standard reference volumes such as Johnson and Kotz (1969,
Chapters 3 and 11). Consider a series of independent trials, in each of
which just one of 3 mutually independent events Al, A2 and A3 must be
observed, and in which the probability of occurrence of event Ah (k = 1 , 2 ,
or 3) is equal to P k for each trial with the sum of the three probabilities
equal to one. The trinomial distribution then is the joint distribution of
 225

the random variables N1, N2 and N3 representing the numbers of

occurrences of the events Al, A2 and A3, respectively, in N trials. It is
defined by

P ( N l ,N,, N,) = M. n (P,

3

k= I
N
INk!)
(6.18)

The distribution of N1, N2 or N3 considered separately is binomial with

P(Nk) satisfying Equation (3.4). Also, if one of the events, say A3, is
ignored, then the other two satisfy Equation (3.4) provided that N is
replaced by N-N3 and Pk by Pk/(l-N3/N). The maximum likelihood
estimator Of PI, (k = 1 , 2 , or 3) is Nk/N.
The preceding theory can be illustrated by means of the following
simple example. For a set of 18 wells on the Canadian Atlantic Margin,
Uuigerina canariensis (Fossil no. 10) and Asterigerina gurichi (Fossil
no. 17) were both observed in N = 5 wells. The exit of no. 10 occurred twice
(N1= 2) above and once (N2 = 1) below that of no. 17, respectively. In the
remaining two wells (N3=2), the events were observed t o be coeval.
Consequently, it can be estimated that no. 10 occurs above, below or coeval
to no. 17 with probabilities of 40,20, and 40 percent, respectively. If coeval
events are ignored, then the estimates of the probabilities that no. 10
occurs above and below no. 17 are 67 and 33 percent, respectively.
Of course, the uncertainties of the preceding estimated percentages
are considerable. For the observed relative frequency 2 6 , the 95 percent
confidence limits for P1 are 5.3 and 85.3 percent, respectively. For 2/3, the
95 percent confidence limits are 9.4 and 99.2 percent. These confidence
limits were looked up in Hald’s (1952) statistical tables. They also can be
computed by using various approximation formulas (see Johnson and
Kotz, 1969, Chapter 3; Southam et al., 1975).
The preceding practical example, clearly illustrates that simple
binomial or trinomial theory results in imprecise estimates of the
probabilities if the sample size N is small. For large samples, however,
the theory is satisfactory.
226
Binomial model based on multiple pairwise comparisons

Gradstein and Agterberg (1982) originally developed the scaling

technique mainly to cope with the problem that nearly all of their sample
sizes (N) were small. Each stratigraphic event Ei was assumed t o occupy a
position along a linear scale L. The positions assumed by Ei in individual
sections fluctuate a t random about an average value along L . The
“distance” D o between average positions of two events Ei and Ej then can
be converted into the probability that Ei occurs above Ej. Alternately, the
probability can be converted into the distance. The advantage of this
method is that the distance need not only be estimated from the relative
positions of Ei and Ej in the sections but (N*) double pairs (Ei, Ek) and (Ej,
Ek) with h f i , j can also be used. Even if all sample sizes are small, N*
may be large, and precise estimates of the average positions of the events
along the linear scale L can be obtained.
Ties were treated as follows. If F(=N1) represents the observed
frequency of A1 (event Ei occurs above Ej), and T ( =N3) the number of ties
for these t w o events, then the score S = N1 + N3l2 was used for
estimating the probability of occurrence of A1 with P1 = SIN. This implies
+
that A2 is observed to occur N2 N3/2 = N-S times, and its probability of
occurrence is P2 = l-S/N. In this approach, a n observed tie receives the
same weight as either one of the direct observations of the events A1 or A2.
However, it is recognized that no preference can be given to A1 or A2 if A3
is observed. Although the average positions of Ei and Ej along the linear
scale could coincide (with P1 = P2 = 0.51, and observed ties will tend t o
decrease the distance between the average positions, the scaling model
does not allow for explicit estimation that two events are observed to be
coeval in a section. Instead of this, a tie is interpreted as a coincidence due
to sampling method (e.g. use of well cuttings) or due to occurrence of
sudden events at the time of deposition which favored fossilization of
several taxa in “patches” (cf. Fig. 2.7).
A distinction should be made between the frequency curve for relative
abundance of occurrence of a fossil taxon through time a t a given place and
the probability curves for its entry and exit (cf. Chapter 2). Methods by
which frequencies are averaged may give range zones which are shorter
than those resulting from “conservative” methods in which more weight is
assigned t o places where events occur relatively high or low in the
stratigraphic column in relation to other events. For example, if exit E l is
observed above exits E2 and E3 in one section but below E2 and E3 in many
 227

other sections, then a conservative method would place the upper limit of
the range for taxon 1 above those of taxa 2 and 3. On the other hand, this
point would fall below those of taxa 2 and 3 when the average location of
E l is determined. From a statistical point of view, the estimation of an
average exit is more satisfactory because the position of the endpoint is
more susceptible t o random fluctuations. Moreover, the average value is
more robust if events are locally out of place due t o anomalous
circumstances such as sediment mixing or misidentification. In the RASC
computer program, individual sections can be compared t o the “standard”
which consists of a set of average distance values along the linear scale L
(normality test; also see Gradstein, 1984).

6.10 Application of Glenn and David’s trinomial model

As outlined in Section 3.4,Glenn and David’s (1960) model is an

extension of the Thurstone-Mosteller model which uses Gaussian curves
for the distribution of positions of events along a linear scale L as is done in

0.4

DIFFERENCE OF EXPECTED VALUES

Fig. 6.13 Probability of a tie as a function of “distance” (6) between mean positions of events along linear
distance scale (after Glenn and David, 1960).
228

the RASC model. As a first step for calculating average distances between
s
events along this scale, the observed “cross-over” fre uencies ( P ) are
converted to 2-values according t o the transformation CP- ( P ) = 2. This is
the inverse of P = Q(2) where 0 denotes the fractile (cumulative
frequency) of a normal distribution in standard form. The model without
ties can be extended to the model with ties as follows.
Suppose that the random variable D represents “distance” along the
linear scale L between two events in a single section. D is assumed to have
unit variance and its average value is 6. Glenn and David (1960) have
introduced a threshold parameter I;. A tie of the two events is assumed to
occur when D is less than T and greater than -I;. The probability of a tie
(P3) then depends on both T and the mean distance (6) between the two
events considered. This relationship is illustrated in Figure 6.13 for T = 0.2
and T = 0.4.
It is readily shown that Glenn and David’s model results in the
following three probabilities for Al, A2 and A3:

P , = D(6-r)

(6.19)

Consequently,

P, + P, = @(6+d

(6.20)

This indicates that 6 and T can be estimated from P I , P2 and P3. A set
of observed frequencies using the format ( F , T I R ) a r e shown i n
Table6.20A. This is the Hay example as used in Agterberg and Nel
(1982b)and Agterberg (1984). It is convenient to define
 229
TABLE 6.20

Example of 10 biostratigraphic events forming optimum sequence as in Agterberg and Nel(1982b, Table
7, p. 74). A. Numbers F , TIR are for pairwise comparison using a trinomial model. If rows are labelled
by the index i and columns by j , then F denotes the number of times EJ follows El in the sequences, T
represents number of ties, and R is number of times E, and E, were observed in the same section.
Example: the first entry of the second column (4,217) indicates that event 1 follows event 2 four times
while the two events were observed to be coeval in two sections. Because R = 7 , this implies that event 1
precedes event 2 in SEQ file for one section. B. Matrix consisting of elements A = ( F + T)IR
corresponding to Table 6.20A.

A
2 1 3 5 7 4 6 8 9 10

2 x 4,217 2,216 3,418 4,116 2,216 2,214 5,015 8,018 4,115

I 1,217 X 1,315 3,318 4,016 4,116 2,113 5,015 8,018 4,115
3 2.2/6 1,315 X 3,216 3,115 3,316 2,214 4,014 6,016 3,114
5 1,418 2,318 1,216 X 3,117 3,317 2,214 5,015 9,019 4,216
7 1,116 2,016 1,115 3,117 X 3,116 1,113 4,115 7,017 4,115
4 2,2/6 1,116 0.316 1,317 2,116 X 2,214 3,014 6,017 3,115
6 0,214 0,113 0,214 0,214 1,113 0,214 X I,012 4,014 1,112
s 0,9/5 0.015 0.014 0.01 5 0,115 1,014 1,012 X 5,OI 5 2,113
Y ri,OlS 0,O/Y 0,016 0.017 1,317 0,014 0,015 X 3,016
lo 0,liS 0,115 0.1 14 0.I15 1,115 0 , I 12 0,113 3,016 X

B
2 1 3 5 7 4 6 9 10

2 x 617 416 414

1 317 X 41 5 313
3 417 415 X 414
5 518 518 316 414
7 216 216 215 21 3
4 416 216 316 414
6 214 113 214 X
8 015 015 014 I12
9 0/8 018 016 014
I0 115 115 114 112

A , . = (F
V V
+ T V )IRV.'

A .. = ( F . .
JI JI
+ TLJ.)IR.,
0 (6.21)

These values are shown in matrix form in Table 6.20B, with A G in the
upper triangle and Aji in the lower triangle. The transformation @-'(AG)
230

was made with the result shown in Table6.21A. Finally, separate

estimates (d and t) of S and T were obtained as

(6.22)

The d-values computed from the values of Table 6.21A are shown in
the upper triangle of Table6.21B and the t-values in its lower triangle.
The d-values can be treated in exactly the same way as the 2-values were
treated in scaling for obtaining average distances between events along
the linear scale L . Each of the t-values can be regarded as an estimate of T.
A frequency distribution of the 32 observed t-values of Table 6.21B is
shown in Table 6.22. Their average amounts to t = 0.4520 which seems t o
be a fairly precise estimate of T. (The standard deviation oft is 0.046.)
Glenn and David (1960) have shown that the preceding simple
averaging method does not result in a least squares solution of T and the
distances between events. They proposed a modified model replacing the
Gaussian curves along the distance scale L by cosine curves. Then the
preceding expressions for d and t represent the least squares solution when
(Aij) is replaced by arcsin(2Aij - 1). Application of the arcsin
transformation to the values of Table 6.20B yields Table 6.21C instead of
Table 6.21A. Table 6.21D was derived from Table 6.21C in the same way
as Table 6.21B from Table 6.21A and also can be used for estimating I; and
the distances. The modified average value now amounts t o t = 0.4080 as
shown in Table 6.22.
A more elaborate test of the preceding version of Glenn and David's
model consisted of its application to 48 events each occurring at least
5 times in the set of 18 wells used by Gradstein (1984). First a n optimum
sequence was obtained (probabilistic ranking followed by modified Hay
method with r n , l = l ) . This sequence was split into t w o segments
consisting of 21 and 27 events, respectively. T was estimated separately by
the two methods (Gaussian Model and Cosine Model) for these two groups
which contain 75 (Group 1 in Table 6.23) and 173 (Group 3) individual
t-values, respectively. Group 2 in Table 6.23 is for 39 t-values arising from
comparison of events in Group 1to events in Group 3. The average values
oft (Gaussian Model) are 0.2419, 0.1914 and 0.2179, respectively. These
 23 1

TABLE 6.21

A. Values CP-1 ( A ) corresponding to Table 6.20B. Values for samples with R = 2 were not used and are
written as x. Values corresponding to 1 and 0 are written as a and -a, respectively. For some subsequent
calculations a was set equal to qc=1.645. B. Values d (in upper triangle) and t (in lower triangle)
obtained by Eq. (6.22). The values aa and aaa are undetermined. C. Same as Table 6.21A except that
the transformation arcsin (2A-1) was used. For some subsequent calculations a was set equal to
qc= 1.571 (instead of 1.645). D. Same as Table 6.21B except that the transformation arcsin (2A-1) was
used.

2 I 3 5 7 4 6 8 9 10

A 2 x 1.068 0.430 1.150 0.967 0.430 a a a a

I -0.180 x 0.841 0.674 0.430 0.967 a a a a
3 0.430 0.891 x 0.967 0.841 a a a a a
5 0.318 0.318 0.000 x 0.180 1.068 a a a a
7 -0.430 -0.430 -0.253 0.180 x 0.430 0.430 a a a
4 0.430 -b.430 0.000 0.180 0.000 x a 0.674 1.068 0.841
6 0.000 -0.430 0.000 0.000 0.430 0.000 x X a X
8 -a -a -a -a -0.841 -0.674 x X a a
9 -a -a -a -a -a -1.068 -a -a x 0.000
10 -0.841 -0.841 -0.674 -0.930 -0.841 -0.253 a -0.430 0.000 x

g 2 x 0.624 0.000 0.416 0.699 0.000 0.823 aaa aaa 1.243

I 0.444 x 0.000 0.178 0.430 0.699 1.038 aaa aaa 1.243
3 0.430 0.841 x 0.484 0.547 0.823 0.823 aaa aaa 1.160
5 0.734 0.496 0.484 x 0.000 0.444 0.823 aaa aaa 1.038
7 0.269 0.000 0.294 0.180 x 0.215 0.000 1.243 aaa 1.243
9 0.430 0.269 0.823 0.624 0.215 x 0.823 0.674 1.068 0.547
6 0.823 0.607 0.823 0.823 0.430 0.823 x X aaa X
8 aa aa aa aa 0.402 0.000 x X aaa 1.038
9 aa aa aa aa aa 0.000 aa aa x 0.000
10 0.402 0.402 0.486 0.607 0.402 0.294 x 0.607 0.000 x

c 2 X 0.796 0.340 0.848 0.730 0.340 a a a a

-0.143 x 0.644 0.524 0.340 0.730 a a a a
0.340 0.644 x 0.730 0.644 a a a a a
0.253 0.253 0.000 X 0.143 0.796 a a a a
-0.340 -0.340 -0.201 0.143 X 0.340 0.340 a a a
4 0.340 -0.340 0.000 0.143 0.000 X 1.571 0.524 0.796 0.644
6 0.000 -0.340 0.000 0.000 0.340 0.000 X X a X

8 -a -a -a -a -0.644 -0.524 X X a a
9 -a -a -a -a -a -0.796 -a -a X 0.000
10 -0.644 -0.644 -0.524 -0.340 -0.644 -0.20 1 X -0.340 0.000 X

~~

D 2 x 0.469 0.000 0.298 0.535 0.000 0.785 aaa aaa I . I07

I 0.326 x 0.000 0.135 0.340 0.535 0.955 aaa aaa I . I07
3 0.340 0.644 X 0.365 0.422 0.785 0.785 aaa aaa I .047
5 0.550 0.388 0.365 X 0.000 0.326 0.785 aaa aaa 0.955
7 0.195 0.000 0.221 0. I 4 3 X 0. I70 0.000 1.107 aaa I . 107
4 0.340 0.195 0.785 0.469 0.170 X 0.785 0.524 0.796 0.422
6 0.785 0.615 0.785 0.785 0.340 0.785 X X aaa X

8 aa aa aa aa 0.464 0.000 X X aaa 0.955

9 aa aa aa aa aa 0.000 aa aa X 0.000
10 0.464 0.464 0.524 0.615 0.464 0.221 X 0.615 0.000 X

values are not significantly different from each other at the 5 percent level
of significance when analysis of variance is applied. This demonstrates
232
TABLE 6.22

Frequency distribution of t-values shown in lower triangles of Tables 6.21B and 6.21D. G.M. denotes
Gaussian Model; C.M. -Cosine Model; N - sample size; S.D. - Standard Deviation.

Class Limits C.M. C.M.

0.000 4
0.001 - 0.200 1
0.201 - 0.400 5
0.101 - 0.600 11
0.601 - 0.800 5
0.801 - 1.000 6

N 32 32
Mean 0.4520 0.4080
S.D. 0.2603 0.2489
S.D./NS 0.0460 0.0440

TABLE 6.23

Glenn and David’s trinomial model applied to 48 exits of Cenozoic Foraminifera observed in 18 wells on
northwestern Atlantic Margin. Abbreviations as in Table 6.22. Groups resulted from splitting the
optimum sequence after 21 events. Group 1 (see Table 6.24 for original data) is for pairwise comparisons
of events belonging to first 21 events, Group 3 is same for last 27 events, and Group 2 is for comparison of
events of Group 1to events of Group 3.

Croup I Group 2 Group 3

Class Limits G.M. C.M. G.M. C.M. C.M. C.M.

0.000 28 28 20 20 73 73
0.001 - 0.200 5 14 1 2 13 21
0.201 - 0.400 22 14 14 8 47 43
0.401 - 0.600 II I1 8 6 27 25
0.601 - 0.800 6 6 3 3 10 10
0.801 - 1.000 I 1 0 0 2 2
1.001 - 1.200 1 0 0 0 1 0

N 75 75 39 39 173 173
Mean 0.2419 0.2242 0.1914 0.1854 0.21 79 0.2008
S.D. 0.2402 0.2321 0.2196 0.2249 0.2288 0.2204
S.D./N% 0.0277 0.0268 0.0352 0.0360 0.0174 0.0168
 , 77
221
10
65
22
17
67
16
71
0 0000
0 3377
D 0642
0 1760
0 0114
0 0082
0

0
1427
0 2832
1832
ELPHlDlUM

COSCINODISCUS SPP
SP
C A S S I D U L I N A TERETIS
UVI CER INA CANAR IENS IS
C O S C I N O D I S C U S SP1

A S T E R I C E R I N A CUR1 C H I
SCAPHOPOD SP1
CERATOEULIMINA C O N T R A R I A
E P I S T O M I N A ELEGANS
233

18 0 145) SPIROPLECTAMMINA CAR1 NATA

21 0 0971 CUTTULINA PROELEMA
20 0 1714 C V R O l O l N A C IRARDANA
15 0 1025 CLOEICERINA PRAEEULLOIOES
26 0 4465 U V I C E R I N A DUMELE I
70 0 0525 A L A E A M I N A WOLTERSTORFFI
24 0 1264 TURR IL I NA ALSATl C A
25 0 0552 COARSE ARENACEOUS SPP
27 0 3457 EPONIOES UMEONATUS
a1 o 0403 C L O E I C E R I N A VENEZUELANA
69 0 1212 NODOSARIA S P 8

rl 33
31
82
29
34
0
0

0
0
0
0718
0334
1809
0862
0123
TUREOROTALIA POMEROLI
PTEROPOD SP1
C L O E I C E R I N A LINAPERTA
C V C L A M U I N A AMPLECTENS
MARGINULINA DECORATA
85 0 1226 PSEUDOHASTICERINA M I C R A
40 0 0062 EULlMlNA ALAZANENSIS
iia o 1178 EPISTOMINA SP5
41 0 1270 P L E C T O F R O N D I C U L A R I A SP1
30 o ogao CIEICIDOIDES ELANPIEDI
35 0 0930 SPIROPLECTAMMINA DENTATA
42 0 0544 C l E l C l D O l D E S ALLEN1
32 o 2685 OUADRIUORPH I N E L L A INCAUTA
86 o oa84 TURR IL I N A B R E W S P IR A
49 0 1135 OSANGULARIA EXPANSA
53 0 1452 U V I C E R I N A EATJESI
57 0 0505 SP IROPLECTAMMINA S P E C T A E I L I S
90 0 0138 A C A R I N I N A DENSA
36 0 0912 PSEUOOHASTI CER IN A W I LCOXENS I S
93 0 0684 A C A R I N I N A AFF BROEOERMANNl
45 0 1372 EULIMINA TRI CONALIS
43 0 1215 E U L l M INA MIDWAVENS I S
50 0 10.91 SUEBOTINA PATAGON IC A
46 0 1150 MEGASPORE S P l
54 0 2800 T E X T U L A R I A PLUMMERAE
52 0 3469 ACAR I N I N A SOLDAOOENS I S

r 56
5s
59
o
0
31137
1139
GLOMOSPIRA CORONA
GAVELI N C L L A BECCAR IIFORM Is
RZEHAKINA EPICONA

n
" .
t
9

I N T E R F O S S I L DISTANCES

Fig. 6.14 Dendrogram for distances between successive events estimated by Glenn and David's trinomial
model assuming Gaussian probability curves for events. Each event (except the last one) is followed by
estimate of distance connecting it to the event immediately below it. These distances were plotted
toward the left and clustered.
234

r 77
228
0 0000
0 k233
ELPHlDlUM SP
C A S S I D U L I N A TERETIS

10 0 0407 UVIGERINA CANARIENSIS

65 0 0873 COSCINODISCUS 5Pl

22 0 0677 C O S C I N O D I S C U S SPP
I 61 0 1k21 SCAPHOPOD SP1

17 0 0809 A S T E R I CER I N A CUR I C H I

I 16 0 IU96 CERATO8ULlMINA CONTRARI A
1 71 0 2101 EP 1 STOM INA ELEGANS

18 0 1058 SPIROPLECTAUMINA CARINATA

r 21

20
0
0
0954
1726
GUTTULINA PROBLEMA

GYRO I D I N A G I RARDANA

:r 15
26
0

0
0420

4655
GLOEICERINA PRAEBULLOIDES

UVIGERINA DUMBLEI

70 0 0265 A L A E A U I N A WOLTFRSTORFFI

24 0 1529 TURRILINA ALSATICA

25 0 0546 COARSE ARENACEOUS SPP

27 0 3387 EPONIOES UUEONATUS
81 0 0230 CLOBICERINA VENEZUELANA

69 0 1 I50 NODOSARIA SP8

1
33 0 1091 TURBOROTALIA POMEROLI

82 0 0048 CLOBICERINA LINAPERTA

31 0 1858 P T E R O P O O SPY

29 0 0320 CYCLAMUINA AMPLECTENS

85 0 0126 PSEUOOHASTIGERINA MICRA

34 0 1503 MARGINULINA DECORATA
4 0 0 0164 BULIMINA ALAZANENSIS

118 0 3025 E P I S T O M I N A SP5

4 1 0 0795 P L E C T O F R O N O I CUL A R I A SPI
30 0 0940 C I B I C I D O I D E S BL A N P I E 0 1
35 0 0985 S P I R O P L E C T A U M I N A DENTATA

42 0 0919 CIBICIDOIDES ALLEN1

32 0 2260 O U A O R I MORPH I N E L L A I N C A U T A
86 0 0730 TUSR 111N A E R E V l S P I R A

49 0 1181 OSANGULARIA EXPANSA

53 0 0767 UVlGER I N A B A T J E S I
57 0 0816 SPIROPLECTAUUI NA SPECTABILI S

36 0 0095 P S E U O O H A S T I CER I N A W1 L C O X E N S I S
90 0 '114 A C A R I N I N A DENSA
93 0 0216 ACARININA AFF BROEDfRMANNl

45 0 1492 BULI M I N A TR I G O N A L I S
43 0 0823 B U L I M I N A MIDWAYENSI S
50 0 0756 SUEBOTI N A P A T A G O N I C A
46 0 1322 U E G A S P O R E SP1
5 1 0 3215 ' E X T U L A R IA P L U M U E R A E
I 4 52 0 2234 ACARININA SOLDADOENS I S
56 0 4460 GLOMOSPIRA CORONA
55 0 08YO GAVELINELL b BE C C A R 11 F O R M I S
59 RZEHAKINA LPIGONA

INTERFOSS I L D l STANCCS

Fig. 6.15 Same as Fig. 6.14 except that cosine-shaped probability curves (instead of Gaussian curves)
were assumed for events. Note that differences between patterns of Figs. 6.14 and 6.15 are small,
indicting that choice of shape of probability curves for events probably is not ofcritical importance.
 235

TABLE 6.24

Estimation of probabilities (PfandP t ) and frequencies (fe and 1,) corresponding to observed successions (fl
and ties ( t ) . Trinomial model was applied to first 21 events (Group 1)of optimum sequence for 48 exits of
Cenozoic Foraminifera also used in Table 6.23 and Fig. 6.14. Last columns show estimated values for
scores (s) based on modified binomial model using RASC weighted distance analysis. See text for
explanations of other column headings. Event numbers of column 1 are explained in Fig. 6.14(from
Agterberg, 1984).

10-17 2.215 0.320.53 2.70.180.9 1.0 0.190.65 3.3 I 8 20 7,5115 0.24 0.50 7.' 0.lY 2.8 Y.I 0.10 0.62 Y.l
10-16 1,116 0.47 0.59 1.5 0.17 1.0 1.5 0.94 0.83 5.0 18-11 10,31 I6 0.41 0.57 9.1 0.18 2.X I1.I 0.14 0.71 11.1
10-17 2.111 0.76 0.70 2.1 0.15 0.4 2.5 1.41 0.92 2.8 I 8 26 7,018 0.12 0.01 4.9 0.16 1.1 7.0 0.68 c1.7> b.'i
17~16 3,116 0.15 0.46 2.8 0.19 1.1 4.5 0.54 0.71 4.2 18-10 6.118 0.96 0.77 6.1 0.12 1.0 6.5 1.01 0.8'4 6 8
17-71 2.011 0.43 0.58 1.7 0.18 0.5 2.0 1.01 0.85 2.1 18 2 6 10.1112 I .02 0.78 9.4 0.12 1.4 10.I 1.16 0.88 10.1
17-18 6.117 0.62 0.61 4 . 5 0.16 1.1 6.5 1.19 0.88 6.2 18 2 I 11.0112 1.14 0.82 9.8 0.10 1.2 11.0 1.41 <>.Y2 11.1
17-20 6.117 C.86 0 . 7 1 5.1 0.13 0.9 6.5 1 - 9 9 0.91 6.5
17-15 6.117 1.03 0.79 1.5 0.11 0.8 6.1 1.71 0.96 6.7 2011 8.2114 0.17 0.47 6.h'O.IY 2.6 9 0 0.2I 0.60 X.1
65 16 4,115 -0.09 0 . 3 7 1.9 0.19 1.0 4.5 0 . 8 4 0.80 9.0 2026 1.118 0.27 0.11 4.1 0.18 1.1 5.5 0.18 Oh5 1.2
77-228 1,214 0.00 0.40 1.6 0.19 0.8 2.0 0.00 0.50 2.0 2070 6.017 0.72 0.68 4 . 8 0.15 1.0 0.0 0.71 0.76 1.1
228-22 2.013 0.58 0.63 1.9 0.16 0.5 2.0 1.21 0.89 2.7 2024 9.1111 0.77 0.70 7.7 0.14 1.6 9.5 0.86 L.81 8.Y
2025 9;1/10 0.YO 0.74 7.4 0.11 1.3 9.3 1.11 0.87 8.7
16-22 3,011 0.230.10 2.10.190.9 3.0 0.01 0.51 2.5 2027 5,017 0.95 0.76 1.3 0.12 0.9 1.0 1.13 0.87 6.1
16-67 3,216 0.140.46 2.80.19 /.I 2.0-0.320.89 2.9
16-71 1,116 0.28 0.12 1.1 0.18 1.1 l . > 0.49 0.69 4.1 I526 4,018 0.10 0.45 3.6 0.19 1.3 l.O 0,lb 0.16 4 4
16-18 Il.1llb 0.47 0.59 9.4 0.17 2 . 8 12.1 0 . 6 4 0.74 11.8 I17U 4.318 0.11 0.62 1.0 0.16 1.1 I.5 0.47 0.68 5.5
16-20 11.2114 0.71 0.68 9 . 1 0.15 2.1 12.0 0.95 0 . 8 3 11.6 I12u 9.1112 0.60 0.64 7.7 0.18 2.1 9.5 q.62 0.73 8.8
16-15 11,4115 0.88 0.74 1 1 . 1 0.11 2.0 11.0 1.18 0 . 8 8 13.2 1525 10.1112 0.73 0.69 8.2 0.15 1.8 10.5 0.8Y 0.81 9.8
16-26 7,018 0.980.77 6.2O.iZ 1.0 7.0 1.310.91 7.3 1527 5.017 0.78 0.71 4.9 0.14 1.0 1.0 0.89 0.81 1.7
1581 4.116 1.13 0.81 4.9 0.19 0.6 4.5 1.30 0.90 1.b
22-71 2.111 0.52 0.61 1.80.170.5 2.) 0.48 0.68 2.0
22-21 2,113 0.840.13 2.20.130.4 2.5 ..800.79 2.4 26 24 5,117 0.10 0.60 4.2 0.17 1.2 5.1 0.08 0.69 '4.8
22-18 1,011 0.700.68 1.4 0 . 1 5 0 . 8 1.0 0.610.74 1.7 26 25 I,014 0.63 0.65 2.6 0.16 0.6 1.0 0.71 0.77 11
22-20 4,115 0.940.76 1.80.120.6 4.5 0.930.82 4.1 26 27 Lli5 0.68 0.61 3.4 0.15 11.8 1.5 0.75 0.77 1.,J
22-15 4.015 1 . 1 1 0.81 4.00.100.5 4.0 1.170.88 4.4 70 2'4 4,016 0.05 0.43 2.6 0.19 1.1 4.0 0.15 0.56 1.4
61-21 U,ll5 0.710.70 3.50.150.7 4.5 0.840.80 4.0 70 25 6,118 0.18 0.48 3 . 8 0.19 I 1 6.1 0.42 0.66 5.1
67-18 1.116 0.61 0.64 1.9 0.16 1.0 5.1 0.68 0.71 4.1 70 27 2.011 0.21 0 . 5 0 1.1 0.19 0.6 2.0 0.U2 0.66 2.0
71-21 1,113 0.3) 0.54 .1.6 0.18 0.5 1.1 0.32 0.63 1.9 70 81 2,011 0.18 0 . 6 3 1.9 0.16 0.1 2.0 0.81 0.80 2.Q
71-18 2.216 0 . 1 8 0 . @ 8 2.90.19 1 . 1 3.C 0.150.56 3.4 70 11 3,014 0.81 0.72 2.9 ,O.I4 11.6 3.0 1.02 rr.81 l.b
71-20 i.I/b 0.410.17 3.40.18 1.1 4.5 0.460.68 4.1 24 2 5 6,119 0.12 0.41 4.1 0.19 1.7 6.5 0.26 0.66 3.4
71-15 4.216 0.600.64 1 . 8 0 . 1 6 1.0 5.0 0.700.76 4.5 24 27 4,016 0.18 0.48 2 . 9 0.19 1.1 '1.0 0.27 0.61 1.6
71-26 4.015 0.700.68 1.4 0 . 1 5 0 . 8 4.0 0.840.80 4.0 2'1 81 3,014 0.11 0.61 2.4 0.17 0.7 3.0 0.83 0.80 3.2
71-27 2.113 1.18 0.87 2.60.080.2 2.3 1.580.94 2.8
2127 1,015 0.06 0.41 2.1 0.19 1.0 3.0 0.'10 0.50 2.1
2148 1.1111 - 0 . 1 50.15 3.8 0.19 2.1 1.5 -0.17 0.43 4.8 2331 7.119 0.61 0.61 1.9 0.16 1.4 7.1 0.60 0.71 6.)
2120 1.619 0.10 0.44 4.0 0.19 l.'7 6.0 0.14 0.15 5.0 2781 2.011 0.31 0.14 1.6 0.18 0.1 2.0 0.41 5.66 2.0
11-15 9.2110 0.27 0.11 5 . 1 0.18 1.8 5.0 0.38 0.65 6.5 2711 5,017 0.18 0.63 4.4 0.16 1.1 1.0 0.60 0.73 5.1
21-26 2,014 0.170.55 2.20.180.7 2.0 0.120.70 2.8 2782 2,011 0.61 0.61 1.9 0.16 0.5 2.0 0.68 0.71 2.1
21-70 >,I16 0.820.72 4.10.14 0.8 5.5 0.810.80 4.8 81 11 l.Oi5 0.21 0.50 2.5 0.19 0.9 1.0 0.19 0.18 2.9
21-24 7.118 0.87 0.74 1.9 0.13 1.1 7.5 1.00 0.84 6.7 8182 1,114 0.27 0.51 2.0 0.19 0.7 3.1 0.27 0.61 2.4
2127 5,016 1.05 0.79 4.7 0.11 0.7 5.0 1.26 0.90 5.4 11-82 4.015 0.01 0.42 2.1 0.19 1.0 4.0 0.08 0.11 2.7

that Glenn and David's trinomial model indeed can be used for describing
the frequencies of observed ties.

The d-values were treated a s 2-values in the RASC computer

program (now setting m,2 = 3 and using the unweighted method for
scaling). The resulting dendrograms are shown in Figure 6.14 (Gaussian
Model) and Figure 6.15 (Cosine Model). It may be concluded that the
differences between results obtained by these two models are minimal. On
average, successive distances in Figures 6.14 and 6.15 are shorter than
those in dendrograms resulting from runs with the RASC program. All
successive distances in Figure 6.14 are less than 0.5. Because T; is
236

TABLE 6.25

Comparison of observed and estimated frequencies for 75 pairwise comparisons of Table 6.24. First six
columns are for trinomial model and last three columns for binomial (RASC weighted scaling) model. If
model provides good tit, the U-values are approximately distributed as chi-squared with single degree of
freedom. Totals are shown in bottom line.

Te To "t Fe Fo 'f e
'
9.09 13 1.69 33.31 39 0.97 45.85 45.5 0.00
10.91 12 0.11 44.53 49 0.45 53.49 53 0.00
9.14 12 0.&9 40.30 41 0.01 45.77 47 0.03
8.90 15 4.18 30.07 29 0.04 35.93 36.5 0.01
10.52 10 0.02 46.72 51 0.39 54.95 56 0.02
8.86 5 1.68 36.00 42 1.00 42.51 44.5 0.09
8.35 6 0.66 34.28 36 0.09 39.58 39 0.01
10.34 4 3.88 32.14 39 1.46 40.39 41 0.01
7.15 2 3.71 22.60 29 1.81 26.29 30 0.52
- -- -- -- --
83.25 79 16.83 319.95 355 6.22 384.76 392.5 0.70

approximately equal to 0.2, most probabilities of a tie between successive

events are about 15 percent (cf. Fig. 6.13).

6.1 1 Comparison of observed and estimated probabilities

A detailed comparison of estimated trinomial and binomial prob-
abilities with observed frequencies is shown in Table 6.24 for Group 1 in
Table 6.23 only. A temporary change in notation restricted to this section
is that f, t , r and s are used instead of F , T , R and S for pairwise compari-
son; F , T and S will be used instead to denote sums of f-, t- and s- values
(see Table 6.25). The distances d f i n Table 6.24 are as in Figure 6.14. For
example, the distance 0.32 between events 10 and 17 is equal t o the sum of
 237

three successive differences (0.0643, 0.1760 and 0.0814) in Figure 6.14.

According t o the original equations for the Glenn-David model, the
estimate of 'c ( = 0.2419) should be subtracted from these distances and the
fractile of the normal distribution in standard form determined for
estimation of PI. In order to distinguish it from another estimate of P1 (see
later), this estimate is written as Pf. For example, the distance
df = 0.3217 gave df-t = 0.3217-0.2419 = 0.0798 from which Pf = 0.53 was
derived. Multiplication of the estimated probability Pfby sample size r = 5
resulted in the estimated frequency fe = 2.7 for number of times event 10
occurs above event 17. This estimated frequency can now be compared t o
the observed frequency f = 2 in the second column of Table 6.24.
It is also possible to estimate P2 and P3. Because P2 = l-P1-P3,the
probability of a tie, written as Pt, is shown only, followed by the
corresponding estimated frequency te. For the previous example,
Pt = 0.18 and te = 2.8 (to be compared t o t=2).
The 75 pairs of events are divided into 9 groups in Table 6.24. The
estimated frequencies te and fe were added for these groups, with the
totals shown as Te and F e in Table 6.25 for comparison to corresponding
sums of observed frequencies written as To and F,. The quantities
Ut = (To- Te)2/Teand Uf = (F, - Fe)2/Fe are also given in Table 6.25.
If the model provides a good fit to the observations, each of the
quantities Ut and Ufis approximately distributed as chi-squared with a
single degree of freedom. The totals C u t and CUf would be distributed as
chi-squared with approximately 9 degrees of freedom. The 95 percent
confidence limit for this distribution amounts t o 16.9. This suggests that
the observed frequencies are well described by Glenn and David's model.
On the other hand, the discrepancy that the Te-values are less than the
To-values in the upper part of Table 6.25 and greater in its lower part may
be significant. The number of degrees of freedom is not known exactly for
this test. It is, however, probably less than 9 and this would increase the
95 percent confidence limit from 16.9 to below C u t = 16.8.
In this chapter the method of scaling was presented and initially
illustrated by using the two examples of the previous chapter on ranking
(Lower Tertiary nannoplankton from the California Coast Range and
Cenozoic Foraminifera from the northwestern Atlantic continental
margin). The basic assumptions of this approach were tested by using
artificial data sets consisting of ranking normal numbers and computer
238

simulation experiments. Important options of the RASC computer

program introduced in this chapter were the normality test, the marker
horizon option and the unique event option. By using the same two
examples it also was shown that a modified version of the trinomial model
of Glenn and David (1960) can be used for description of observed frequen-
cies of coeval biostratigraphic events. The stratigraphic significance of the
threshold parameter T is not immediately obvious. It can be said that a
new distribution for ties (see Fig. 6.13) has been introduced in addition to
the probability distributions for events along the linear scale L. The
height of the new distribution for ties is roughly proportional to the value
of 7;. In general, T therefore expresses the likelihood that events are coeval.
In the RASC model, observed ties are not ignored but each tie of two
events Ei and Ej is scored as a 50 percent probability that Ei occurs above
Ej and a 50percent probability that Ej occurs above Ei. The last four
columns of Table 6.24 show observed scores in comparison with estimated
frequencies. The estimated probabilities P, (for Ei occurring above Ej)
satisfy P, = W d , ) where d, was estimated by means of the weighted
scaling option of the RASC computer program in which variations of
sample size are considered. The agreement between observed and
estimated scores is excellent (also see Table 6.25, for comparisons of group
totals, S, and So for estimated and observed scores, respectively). Because
the origin of the RASC scale is set at the location of the first event in a
scaled optimum sequence, N events obtain N* ( = N - l )cumulative RASC
distances after scaling. In general, these N* values can be used t o
estimate the N(N-1)/2probabilities that one event occurs above (or below)
another event. These expected probabilities for pairwise comparison are
close t o the observed probabilities, because the former were computed from
the latter. This conclusion is supported by application of the chi-squared
test for goodness of fit after grouping pairs of events (cf. Table 6.25, last
column). The number of degrees of freedom to be used in this chi-squared
test, however, remains unknown, because of autocorrelation of the
estimated RASC distances. The latter topic will be discussed in more
detail in Chapter 8 in relation t o the normality test.
In addition to providing a good fit, the RASC method has several
options (normality test; marker horizon, unique event and weighted
scaling options) which are not available for the modified Glenn-David
model. For these reasons, this trinomial model should only be used when
it is necessary t o model observed frequencies of coeval events.
 239

CHAPTER 7
RANK CORRELATION AND PRECISION OF SCALED OPTIMUM
SEQUENCE

7.1 Introduction
Suppose that a number of objects has been ranked in two different
ways, e.g. by using different characteristics. One then may be interested
in the mutual agreement or disagreement of the two rankings. Rank
correlation methods are described in detail by Kendall (1975). Many
authors have applied these methods in biostratigraphy for comparing
sequences of events, e.g. as obtained by different methods, with one
another (see, for example, Brower, 1985,1989; Harper, 1984). In the first
part of this chapter, rank correlation will be discussed in connection with
the RASC step model. Examples of application will be given. A method for
estimating the precision of the cumulative RASC distances of the scaled
optimum sequence will be presented in the second part of this chapter.

7.2 Rank correlation coefficients

The two measures of rank correlation discussed by Kendall and
widely available in systems of statistical software (e.g. SAS) are Kendall’s
and Spearman’s rho (p). They are estimated by using the following
tau (T;)
equations:

(7.2)

+
where S is a total score of 1for pairs of elements having the same order
in both series and -1otherwise. The total number of elements is written as
n. Spearman’s rho is based on the sum of squared differences (SSD) of
rankings of the elements in the two series compared t o one another.
240

Both rank correlation coefficients emulate Pearson’s product-moment

correlation coefficient for a bivariate relationship in that they vary
+
between 0 for lack of correlation and 1 or -1 for maximum positive or
negative correlation. Unless there is complete a g r e e m e n t o r
disagreement, tau and rho are not the same for any given pair of rankings.
Rho tends to give more weight t o inversions of ranks which are farther
apart. In practice, it is often found that, when neither coefficient is close to
unity, rho is about 50 per cent greater than tau in absolute value (Kendall,
1975, p. 12). Although rho is easier to calculate than tau, Kendall has
shown that from practical as well as theoretical points of view, tau is
preferable to rho. For example, after completing two rankings of the same
set of objects, it may be that some new objects become available for
ranking. In that situation, rho must be completely recalculated, whereas
the addition of new members does not require a complete recalculation of
tau. For the latter reason, it is also easier to evaluate the influence of
addition of individual objects on tau than on rho.
Kendall’s (1975, p. 3) first example consists of the following two
rankings of ten objects A, ...,J:

A B C D E F G H I J
Rankingl: 7 4 3 10 6 2 9 8 1 5
Ranking2: 5 7 3 10 1 9 6 2 8 4

Then n = 1 0 objects have n(n-1)/2=45 possible pairs. Table 7.1 is a

complete list of scores being +1 if two elements forming a pair have the
same order in both rankings; and -1, otherwise. In total, there are P=21
positive and Q = 24 negative scores in this table. The sum of all elements
is equal to -3. Hence, according to Equation (7.11, t = -0.07.
In order to estimate Spearman’s rho, the sum of squared differences
(SSD) is needed. Individual squared differences for the examples are
shown in the following tabulation:
 241
TABLE 7.1

Listing of all 45 pairs and their scores for Kendall’s(1975) first example with 10 rank members A-J.

Pair Score Pair Score

AB -1 CJ +1
AC +1 DE +1
AD +1 DF +1
AE +1 DG +1

AF -1 DH +1
AG +l DI +1
AH -1 DJ +1
A1 -1 EF -1

AJ +l EG +1
BC +1 EH +1

BD +l EI -1
BE -1 EJ -1
BF -1 FG -1
BG -1 FH -1
BH -1 FI +l
BI -1 FJ -1
BJ -1 GH +1

CD +1 GI -1
CE -1 GJ +1

CF -1 HI -1
CG +1 HJ -1
CH -1 IJ -1
CI -1

A B C D E F G H I J
Ranking 1: 7 4 3 1 0 6 2 9 8 1 5
Ranking 2: 5 7 3 101 9 6 2 8 4
242
Differencesd 2 -3 0 0 5 -7 3 6 -7 1
Differences2d2 4 9 0 0 25 49 9 36 49 1

By summing the entries i n the bottom row, we find S S D = 1 8 2 .

Consequently, according t o Equation (7.2) p = -0.103 which is somewhat
smaller than t = -0.07.
Kendall’s tau and Spearman’s rho have been calculated for the optimum
sequences of Table 6.9 obtained by running RASC on 50 artificial
sequences in computer simulation experiments. Table 7.2 shows the two
ranking correlation coefficients between every optimum sequence and the
underlying true sequence consisting of integer numbers from 1 to 20. All
ranking statistics of Table 7.2 are rather large, indicating relatively
strong positive correlation. As expected, there is a general decrease in
strength of correlation when the spacing between expected values along
the real line decreases from 1.0 to 0.1. For set 1, scaled optimum sequences
are somewhat better than optimum sequences obtained by ranking but the
opposite holds true for set 2. From these computer simulation
experiments, it cannot be decided which type of optimum sequence is best.
It only can be concluded that these optimum sequences are approximately
equally good. A similar conclusion will be drawn from the results of
Harper’s (1984) computer simulation experiments to be discussed in
Section 7.4. It does not follow from this conclusion that ranking of
stratigraphic events is t o be preferred to scaling because the latter
technique requires more computing. In practical applications, the
advantage of scaling with respect to ranking is that clusters of events
separated by hiatuses can be identified so that a regional biozonation can
be constructed. It is desirable that the optimum sequence obtained by
ranking which forms the input for scaling is as good as possible because
estimates of intervals between successive events are less precise if the
events subjected to scaling are out of order (cf. Section 7.5).

7.3 RASC step model

In RASC, stratigraphic events are assigned numbers in the dictionary
and these numbers are used in the rankings. Suppose that the 10 objects
(A, ...,J) of Kendall’s first example are numbered 1 to 10:
 243

TABLE 7.2

Kendall’s tau and Spearman’s rho for optimum sequences of Table 6.9 correlated to underlying true
sequence consistingof integer numbers from 1to 20.

A(Set 1) Tau Rho B(Set2) Tau Rho

Ila-e 0.990 0.999 IIIc-e 0.979 0997
IIIa-e 0.979 0.997 IVa-b 0.979 0.997
IVa-b 0.947 0.990 IVC 0.958 0.994
IVC 0.947 0.991 IVd-e 0.968 0.996
IVd-e 0.968 0.994 Va 0.895 0.979
Va-b 0.853 0.955 Vb 0.884 0.970
vc 0.884 0.974 vc 0.863 0.961
Vd-e 0.874 0.970 Vd-e 0.863 0.959

1 2 3 4 5 6 7 8 9 10

Rankingl: 7 4 3 10 6 2 9 8 1 5

Ranking2: 5 7 3 10 1 9 6 2 8 4

Then the rankings rewritten as RASC input sequences become:

Sequencel: 9 6 3 2 10 5 1 8 7 4

Sequence2: 5 8 3 10 1 7 2 9 6 4

In the RASC step model, which can be applied after computation of an

optimum sequence, the observed sequences for all stratigraphic sections
are compared with this optimum sequence. The latter represents a n
average ranking based on the observed sequences for all sections. Suppose
that, in Kendall’s first example, sequence 1 is the optimum sequence and
sequence 2 is one of many section sequences on which sequence 1 is based.
In the step model, the position of each event is compared to its position in
the optimum sequence. A penalty point is scored each time the event is out
of place with respect to another event in comparison with their order in the
244
optimum sequence. Table 7.3 shows the penalty points scored for the
example.
Table 7.3 has separate columns for the number of times an event
occurs “too high” or “too low” in the section. For example, event no. 9, with
position no. 1 in the optimum sequence, occurs three places from the
bottom in the section. It occurs “too high” with respect to all other events
in the optimum sequence except events 6 and 4. Its total number of
penalty points is equal to 7. Another example is as follows. In the section,
event no. 1occurs above nos. 2,9and 6,instead of above these events as in
the optimum sequence. Consequently, it has penalty score 3 for occurring
“too high”. Its other penalty point arises because, in the section, event 1 is
observed below event 8. Event 1’s total score, therefore, is 4 penalty
points. The column totals for “too high” and “too low” must be equal to one
another. It also can be seen that these totals are equal t o Q ( = 24 for the
example), representing the total number of -1 scores used previously for
estimating S, which is needed t o compute tau (see Table 7.1). P can be

TABLE7.3

Comparison of assignment of penalty points in RASC method with computation of t a u on basis of

Kendall’s first example. Sums of columns for events that a r e “too high” and “too low” a r e both equal to
Q=24. Total number of penalty points is 2Q=48. Tau is fully determined by Q and total number of
events ( n = 10).

Optimum Event in Event in

Optimum Section Penalty
Sequence “too low” Points
Position Sequence (Sequence 2) “too high”

9 5 0 7 7
1
6 a 0 7 7
2
3 3 2 2 4
3
2 10 2 5 7
4
10 1 3 2 5
5
5 7 5 0 5
6
1 2 3 1 4
I
a 9 6 0 6
8
7 6 3 0 3
9
4 4 0 0 0
10
Sum = 24 24 48
 245

obtained from Q because P + Q=n(n-1)/2,representing the total number of

ordered pairs of events.
Suppose that the total number of penalty points is written as T ( =
2Q). Then the relation between T and T can be written as:

(7.3)

This equation, for example, can be used to evaluate the relative strength of
correlation of each of the.,three series in the previous example of Table
6.11. It already was pointed out that the total numbers of penalty points
amount to 22, 33 and 28 for the situations of Tables 6.11A, B and C,
respectively. Because n=20, it follows from Equation (7.3) that the
corresponding tau-values are 0.884,0.826 and 0.853.
Table 7.4 shows another example of application. The 25 original
input sequences of Table 4.15 (cf. Sections 4.9 and 6.5) were correlated to
the scaled optimum sequence extracted from this dataset after final
reordering (see Fig. 6.10). All tau-values for rank correlation in Table 7.4

TABLE 7.4

Kendall’s tau for 25 sequences of Table 4.15 correlated to scaled optimum sequence of Fig. 6.10. Values
probably different from zero are marked by one (a= 0.05)and two (a = 0.01) asterisks, respectively.

Sea. Tau Seq . Tau

1 0.31* 14 0.39**
2 0.07 15 0.54**
3 0.61** 16 0.26
4 0.44** 17 0.27
5 0.33* 18 0.09
6 0.32* 19 0.37*
7 0.17 20 0.45**
8 0.49** 21 0.57**
9 0.34* 22 0.56**
10 0.42** 23 0.48**
11 0.93** 24 0.03
12 0.40** 25 0.41**
13 0.49**
246

are positive but the differences between values are relatively large. The
smallest tau-value is 0.03 and the largest one is 0.61. Values that differ
significantly from 0 are marked by asterisks in Table 7.4. A single
asterisk indicates that a value exceeds the threshold value for level of
significance equal to a = 0.05; two asterisks mean that the significance
level for a = 0.01 is exceeded as well. Most computer programs for rank
correlation provide statistics for testing the significance of Kendall’s tau
and Spearman’s rho (also see Kendall, 1975, Chapter 4). It can be shown
that S in Equation (7.1)has variance equal to

uarS = n ( n - l ) ( 2 n + 5 ) / 1 8
(7.4)

In the example of Table 7.4, n = 25. Consequently, var S = 1833.3 with

corresponding standard deviation 6 ( S )= 4 2 . 8 2 . For n > 1 3 , S i s
approximately normally distributed. If there is no rank correlation,
E(S)= 0. Then it is possible t o estimate X,representing the smallest value
of S which is significantly different from zero. After application of a
continuity correction (cf. Kendall, 1975, p. 54) which simply consists of
subtracting 1 from X, it follows that

(7.5)

If the absolute value of S is tested, a = 2(1-Pc). If a = 0.05, P, = 0.975 and

Z,=1.96. For the example, 6 ( S ) = 4 2 . 8 2 and Equation (7.5) gives
X=1.96X42.82+1=84.93.
From Equation (7.1) it follows that, for a=0.05, the critical value of
tau is 0.283. If a=0.01, this threshold value becomes 0.372. For this
reason, values in Table 7.4 which are greater than 0.283 and 0.372 are
followed by one and two asterisks, respectively.

7.4 Presorting and ranking by Harper

In a study evaluating various ranking techniques, Harper (1984)
found that probabilistic ranking (presorting option) provided slightly
better rankings than the modified Hay method. Harper was interested in
comparing competing ranking algorithms in stratigraphic paleontology on
the basis of computer-simulated sections. By means of a computer
 247

program he (1) generated a hypothetical, and thus known, succession of

taxa in time, and (2) simulated their succession in strata at several local
sample sites.
If desired, steps (1)and (2)may be repeated for many (50 or 100, for
example) iterations and the local site data for each iteration sent t o user
routines for inferred rankings (inferred succession of events in time).
First, data for first and last occurrences (entries and exits) taken together,
then data for exits-only, then data for entries- only were sent. For each
simulated data set, Kendall and Spearman rank correlation coefficients
were computed, and the inferred rankings compared with the known
succession of events in time. The performance of two competing ranking
algorithms may be compared by

(1) obtaining for each submitted dataset the differences between

corresponding Kendall and Spearman rank correlation coefficients
computed for the two algorithms, and

(2) testing the observed differences for statistical significance.

Harper (1984) used his computer program to compare three ranking
algorithms (presorting, ranking and scaling) provided by Agterberg and
Nel(1982a, b) as well as to determine whether the algoithms work as well
for datasets combining exits and entries versus datasets for exits-only or
entries-only. He concluded from a series of experiments that Agterberg
and Nel’s presorting algorithm ( = probabilistic ranking) performed
somewhat better than the modified Hay and scaling algorithms. All three
methods performed slightly but significantly better on data for exits-only
or entries-only as opposed t o combined data. The reader is referred to
Harper (1984) for a full discussion of his approach and complete results for
all experiments performed. Only a few examples will be given here with
emphasis on how Harper’s approach can be used in practice; e.g.for
choosing the threshold parameters h, and mcl.
The computer program begins by generating ranges for 50 taxa over
80 time intervals. A random number generator is used for determining
“true” entries and exits of each taxon in a range chart. Next stratigraphic
succession data for n, sample sites are generated by random sampling of
the range chart. This sampling is controlled by choosing a value for

(1) the probability ( P I )that a given taxon is present at a local site;

248

(2) the probability (P,) that a taxon is sampled at a given horizon a t a

sample site given that it occurs in the time interval represented by
the horizon; and

(3) the probability (P,) that two adjacent horizons correspond t o the same
time interval.
Harper conducted 3 experiments (A, B and C)of which the parameters are
shown in Table 7.5. For each sample site, nt sets of stratigraphic
succession data were obtained, with nt representing the number of
iterations. Run, sample site, and sequence data were sent to the RASC
computer program in order t o obtain three types of optimum sequences
(a)probabilistic ranking (presorting only); (b) modified Hay method only;
and (c)scaled optimum sequence as derived from (b). The threshold
parameters employed are shown in Table 7.5. Harper (1984, Fig. 4-6)
compared experimentally-obtained optimum sequences with the “true”
optimum sequence on the range chart by using Kendall’s rank correlation
coefficients. In total, 1950 tau-values were calculated, one for each
+
comparison; all turned out to be relatively close to 1, and significantly
greater than zero. This signifies that all rankings were good. However, by
comparing methods with one another, and looking a t small differences
between average tau-values, it can be determined which one of a pair of
techniques is better. Average differences between tau-values for
comparing presorting with the modified Hay method are shown in the
bottom four rows of Table 7.5. Each of the values shown is the average of
50differences between tau-values, except the two values in the last
column which were based on 100 differences; n.0. indicates that an average
for 100 runs was not obtained for Run C. A negative value signifies that
the modified Hay method gave poorer rankings than presorting. Except
for Run B (first run), the negative values are significantly different from
zero as determined by Student’s t-test (Harper, 1984, Tables 2-7). The
results for exits and entries are similar as can be expected, and the first
two values in the last two columns also duplicate one another.

It may be concluded that, for the experiments performed, probabilistic

ranking gave better results than use of the modified Hay method only,
when k, is relatively small. When h, is large, the two methods probably
give rankings that are equally good. The results of the experiments also
suggest the possibility that, by increasing the ratio h,lrn,, the performance
of the modified Hay method can be improved. The presorting option
(renamed probabilistic ranking in Section 5.5) was introduced i n
 249

T A B L E 7.5

Results for three computer simulation e x p e r i m e n t s ( A , B a n d C) c o n d u c t e d by H a r p e r (1984)(for

explanation see text).

A B C
Number of sites: ns 22 16 6
Probability of presence: p, 0.20 0.20 0.10
Sampling probability: p2 0.55 0.80 ax5
Adjacency probability: p3 0.10 0.10 0.20
Number of datasets: nt 50(or 100) 50(or 100) 50
Minimum number of sites: kc 5 7 3
Minimum number of pairs: mc 4 5 3
Ratio: kJm, 1.25 1.40 1 .00
exits - 0.013 - 0.003 - 0.022
Average difference entries - 0.014 - 0.003 - 0.020
between tau-values: both - 0.004 - 0.001 - 0.007
both( 100) - 0.005 - 0.000 n.o.

Agterberg and Nel(1983a) and routinely has been used in RASC runs
after 1980. The results of presorting are independent of the choice of the
threshold parameters m,, and mc2 which apply t o the modified Hay
method and scaling, respectively. As a result of Harper’s experiments, the
RASC program was modified in 1983 to allow the choice of separate
threshold parameters for these two techniques. Before then, all runs
including those performed by Harper had m,, = mC2.
Application of the modified Hay method after probabilistic ranking
can be regarded as a fine-tuning operation in situations when there are
many missing data. The presorting could yield poor results when many
frequencies are undetermined. Then it should be useful to compare the
ranking of each event with all others in order t o find the optimum
permutation as is done in the modified Hay method. Ideally, the threshold
parameter m,, should be set equal to 1 so t h a t all frequencies are
considered. However, a decrease in mCl frequently corresponds to an
increase in number of cycles (inconsistencies involving 3 or more events).
It then is necessary to use a value greater than 1 in order t o reduce the
number of iterations.
Harper (1984) also found negative differences between tau-values
when the modified optimum sequence resulting from scaling was
compared to the optimum sequence resulting from the modified Hay
method only. However, the lower tau-values in this instance may have
been caused by the fact that Harper (1984, p. 16) regarded a s tied
successive events which were less than 0.5 apart along the RASC scale. A
modified formula for estimating Kendall’s rank correlation coefficient was
used t o accommodate tied events. On average, events preceding other
events along the RASC scale, occur before those other events on the range
chart as well, even when distances between successive events are small.
Scoring them as tied, therefore, results in a somewhat smaller tau-value.
This may explain why the optimum sequence from the modified Hay
method, in which no ties were allowed, yielded somewhat higher tau-
values.
Finally, Harper (1984)showed that exits and entries, run separately,
gave somewhat higher tau-values than when both were mixed together.
This was t o be expected (also see Edwards and Beaver, 1978) because, on
the average, exits will be moved downward, and entries upward, with
respect to their relative positions on the range chart when stratigraphic
succession data for sample sites are generated using probabilities of
occurrence (PI,P, and P J . If exits or entries are considered on their own,
this bias will not show up. However, if they are mixed, some exits will
probably assume final positions, in any type of optimum sequence, below
entries of other taxa which occur above them on the range chart. Although
smaller tau-values are t o be expected for sequences of mixed entries and
exits, these differences were almost negligibly small in the results of
Harper’s experiments. Harper’s experiments were limited t o a single type
of artificial dataset. It may be expected that different specific conclusions
would result from other datasets. Nevertheless, the preceding discussions
illustrated that valid generalizations can be derived from computer
simulation experiments.

7.5 Precision of the scaled optimum sequence

On the basis of computer simulation experiments, it was concluded in
Section 6.5 that, in general, it is possible t o obtain unbiased estimates of
the cumulative RASC distances of the scaled optimum sequence, provided
that the order of events in the scaled optimum sequence is close to the true
order of the events. On the other hand, it was not possible t o obtain
unbiased estimates of the standard deviations of the intervals between
successive events along the relative time scale used for the scaling. It was
pointed out (cf. Eq. 6.17)that the indirect distances used for estimating
each interval are not stochastically independent. Consequently, it would
not be a promising approach to add biased variances for the intervals in
 251

order to estimate precision of any cumulative RASC distance which is the

sum of many intervals. It will be shown in this section that, in general, the
jackknife method can be used t o obtain approximately unbiased estimates
of the standard deviations of the cumulative RASC distances if the order of
events in the scaled optimum sequence is close to the true order of the
events. The mathematical background of the jackknife method will be
given in Chapter 10. Here the purpose of this procedure will be discussed
in qualitative terms only, using two of the abbreviated computer
simulation experiments for example.
Table 7.6 shows the complete matrix of 2-values which led to the
scaled optimum sequence of Figure 6.10. It should be remembered that in
this experiment, there are 25 sequences for 20 events which, in each
sequence, occupy values that are 0.1 units apart. The standard deviation
which controls the scatter of individual events about their means is 0.7071
for all events. Because total distance between the expected location of
events 1 and 20 is only (1.9X0.7071=) 1.34 standard deviations for the
difference between two events, none of the 20 events is likely to occur
before of after one or more of the other events in all sections. This explains
why qc = 2.054 does not occur as a 2-value in Table 7.6. The largest 2-
value for this experiment is 1.751 corresponding to P = 0.96, representing
the situation that event 1 occurs before event 19 in 24 of the 25 sequences.
Consequently, it is not necessary to make adjustments for truncation
effects when distances between events are estimated from the 2-matrix
and the following slightly different procedure can be followed.
The bottom row of Table 7.6 shows the average 2-value for each
column. Each column average is based on 19 separate 2-values because
the diagonal elements were not used. These averages can be regarded as
estimates of the expected locations E(X) of the events along the RASC
scale. The origin is between events 11 and 16. If this origin is moved t o
the first event of the scaled optimum sequence by adding 0.709, the RASC
distances of the first column of Table 7.7 are obtained. These values are
approximately equal to the unweighted linear scaling values ( X o ) for this
experiment which are listed in the second column of Table 7.7. The slight
differences between the values in the first two columns of Table 7.7 are due
t o the fact that direct distance estimates are weighted twice as much as
indirect distance estimates when the procedure of Table 7.6 is followed. It
was already noted (cf. Section 6.4) that doubling the weights of direct
distance estimates gives slightly better results. As a procedure it is also
TABLE7.6

Matrix of 2-values of computer simulation experiment of Tables 4.15, Fig. 6.10 and Table 7.4. The 20 events in 25 sequences have expected
values which are closely spaced (at 0.1 intervals) along the RASC scale. The column averages provide estimates of these mean positions
variant of unweighted scaling method, see text for further explanation). Successive values within any column are stochastically independent
because they deviate randomly from their mean values. The latter are for distances from the mean position of the event labelling the column.
The standard deviation of the column average, therefore, can be estimated, e.g. by the jackknife method, without distortion by autocorrelation
effects. This property is preserved when the jacknife method is applied to unweighted or weighted distance estimation a s in the RASC
computer program.

4 3 1 5 6 2 7 9 11 16 14 10 12 15 13 17 18 19 20
8 --
4 x - 151 0 253 0 151 0.253 0.468 0 253 0.468 0468 1175 0 842 0.706 0706 0842 0 842 0994 1405 1405 1405 0 994

3 0 151 x 0.151 - 253 0.151 0.151 0 358 0.253 0.253 0.253 0 842 0.842 0.706 0 842 0.994 0583 1405 0842 1175 1405

I - 253 151 -253 -.050 0 151 0.151 0468 0358 0.106 0 583 0.468 0706 0583 0.842 0994 0 994 0 842 1751 1405

5 - 151 0 253 0.253 I 0 050 0 358 0.253 0.468 0 583 0.468 0 253 0358 0.468 0.583 0 583 0583 0842 0842 1175 1175

6 - 253 -.151 0.050 - 050 - 151 0 151 0 151 0.358 0253 0 994 0583 0.583 0706 0.106 0706 1405 1405 0 994 1175

2 -468 - 151 ~.151 - 358 0.151 x -.050 0.253 0.151 0 358 0 468 0.706 0.583 0.358 0.583 0842 0706 0994 0 994 1405

7 ~ 253 -.358 -.I51 453 -.I51 0.050 ,050 0.358 0050 0 358 0.583 0.706 0.468 0.253 0583 0842 0706 0 994 1175

8 468 - 253 -.468 ~.468 -.I51 -253 0 050 x 0.358 0.253 0 253 0.253 0.358 0.468 0.358 0468 0706 0706 0 842 1405

9 -468 - 253 - 358 - 583 -.358 -.I51 - 358 - 358 x -.358 0 050 0 151 0.358 0.253 0 253 0583 0358 0583 0 468 0 842
II - 1 18 253 - 106 -.468 -.253 - 358 -.050 -253 0.358 x 0 151 -.050 -.050 0.253 0.468 0358 0583 0253 0 468 0 583

16 -842 - 842 - 583 2.53 -.994 -.468 -.358 2.53 -050 - 151 x 253 0.253 0.358 0.358 050 0842 0706 0 706 0 583
I4 - 706 - 842 -.468 -.358 -.583 -.’I06 -.583 -253 - 151 0050 0 253 x -.050 -.050 0.050 0253 0468 0583 0 706 0 994

10 706 - 706 - 706 -.468 -.583 -.583 - 706 -.358 -.358 0.050 - 253 0.050 x 0.151 0.253 0358 0583 0358 0 151 0 706

12 -.a42 -.842 583 -.583 -.706 - 358 - 468 -.468 253 -253 358 0.050 -151 x 0.358 0050 0151 0 151 0 468 0 994
15 - 842 994 442 - 583 - 706 - 583 253 - 358 -253 -.468 358 -.050 -.253 -.358 x 0253 0050 0 253 0 253 0 994
13 -.994 - 583 - 994 - 583 - 706 - 842 -.583 -468 - 583 - 358 0 050 -253 -358 -.050 -253 x 0 151 0253 0 050 0 706

17 -1 41 -1 41 - 994 - 842 - 1 41 -.I06 - 842 -.I06 -.358 -.583 842 -468 ~583 -151 -.050 151 x 0 151 0 358 0 706

18 -1 41 - 842 842 - 842 ~1 41 994 - 706 -.I06 - 583 - 253 706 583 -358 - 151 -.253 253 151 x 0 050 0 706
19 -141 -1 18 - I 75 - 1 18 994 -.994 ~ 994 -842 -.468 -468 - 706 -706 - 151 -468 - 253 050 358 050 0 468

20 - 994 -1 41 -1 41 -1 18 - 1 18 -1 41 -1 18 -1 41 - 842 -.583 - 583 -.994 -706 -994 - 994 -706 706 706 - 468 x

Ave -709 -584 -528 ~473 ~482 -.345 -.277 -.184 -005 0093 0.128 0 140 0.215 0260 0318 0.381 0.599 0591 0.707 0.971
 253

TABLE7.7

Comparison of four scaling methods applied to example of Table 7.6. Ave represents column average of
Table 7.6 after addition of 0.709 (=minus first column average). X,and X are RASC computer program
unweighted and weighted scaling results. E (X)represents true mean value which is multiple of 0.0707.
Q and s ( Q ) are jackknife estimate and jackknife standard deviation using RASC weighted scaling
method. t (X)is studentized deviation of X from true mean value. Penalty points (pp) for event numbers
of column 1are shown in last column.

4 0 000 0 000 0 000 0 212 0 000 0 000 *** 3

3 0 125 0 I33 0 117 0 141 0 170 0 057 -.429 1

1 0 181 0 I68 0 172 0 000 0 179 0 040 4298** 2

5 0 236 0 228 0 185 0 283 0 200 0 064 -1.53 1

6 0 227 0 214 0 204 0 354 0 215 0 052 .2.88* 1

2 0 365 0 340 0 306 0 071 0 319 0 049 4821*' 4

7 0 433 0 420 0 375 0 424 0 417 0 054 -.920 0

8 0 525 0 501 0 453 0 495 0 488 0 054 -.781 0

9 0 705 0 680 0 634 0 566 0 677 0 067 1.019 0

11 0 803 0 741 0 663 0 707 0 636 0 067 ~.651 1

16 0 838 0 793 0 726 I061 0 727 0 059 -5 66** 5

14 0 849 0 812 0 736 0 919 0 774 0 036 -5.14** 2
10 0 924 0 887 0 803 0 636 0 837 0 048 3.499" 3
12 0 970 0 925 0 851 0 778 0 890 0 053 -1.39 2
15 1 027 0 983 0 923 0 990 0 972 0 059 -1.12 0

13 1090 I 083 0 986 0 849 1019 0 057 2.441' 3

17 1308 1234 11.54 1131 1170 0 057 0394 0

18 1300 1226 I170 1202 1188 0 056 0.578 0

19 1417 1343 1265 1273 I281 0 065 0.124 0

20 1 680 I628 IS98 1344 1644 0 063 4.072'' 0

invoked in weighted distance estimation option of the RASC computer

program.
The weighted scaling values ( X ) previously used for constructing the
diagram of Figure 6.10 (also see Table 6.14) are shown in the third column
of Table 7.7 in comparison with the theoretical mean positions E(X).
Jackknife estimates ( Q ) for weighted scaling are presented in the next
column. If the jackknife estimates ( Q ) are close to the weighted scaling
values (X),their standard deviations can be used as standard deviations of
X.
In general, the jackknife provides a non-parametric method of
estimating the mean and its standard deviation for a sample of n
254

independent and identically distributed random variables. In the

situation of ungrouped data, each of the n values is successively deleted
from the sample and a pseudovalue is computed from each reduced data set
with (n-1) values. The jackknife estimate is the mean of the n
pseudovalues. In the situation of Table 7.6, each column average is based
on n ( =19) values for separate events. These values can be regarded as
realizations of stochastically independent random variables for individual
events. Every event corresponds to a set of 25 random normal numbers
with its own mean value. Deletion of an event results in a reduced 2-
matrix without the row and column of the deleted event. The 2-values for
the remaining n-1 (= 18) events are not changed by the process of deleting
a n event. The 19 pseudovalues are not necessarily stochastically
independent but this hypothesis can be tested in the computer simulation
experiment because all deviations from the true means are known.
Studentized residuals t ( X ) were obtained by dividing each difference X -
E(X) by its corresponding standard deviation s(Q)(see Table 7.7).
The 20 studentized residuals of Table 7.7 should have zero mean and
deviate from zero according t o the t-distribution with n-1 ( = 18) degrees of
freedom. Consequently, it would be expected that, on average, only 1 out
of 20 values, in absolute value, deviates by more than 2.101 from zero, and
1 in 100 values by more than 2.878. Most of the studentized residuals in
Table 7.7 are within these confidence levels of Student’s t-distribution for
18 degrees of freedom. However, a number of the studentized residuals are
too large in absolute value indicating that locally the hypothesis of
stochastical independence of the pseudovalues was not satisfied.

One problem here is t h a t the origin of a RASC scale is set

automatically at the first event of the scaled optimum sequence. All
pseudovalues are forced to be zero a t this point and this results in the
artificial result s(Q) = 0 for first events. This problem generally cannot be
avoided in practical applications. Another problem indicated by the
results of Table 6.15 is that anomalously large values occur a t positions in
the scaled optimum sequence for events that are out of position with
respect to the true squence of expected values. The last column in Table
7.7 shows number of penalty points for each event. For example, event 16
ended up in position 11 of the scaled optimum sequence. For this reason, it
was assigned (16-11=) 5 penalty points. Its studentized residual ( = -5.66)
is nearly twice as large as the significance limit ( = 2.878) for a = 0.01. This
suggests that s(Q)( = 0.059) for this event is too small by a factor of two or
more. It is noted that the jackknife procedure applied t o standard
deviations obtained by means of Equation (6.13) does not remove bias from
these estimates as illustrated in Table 7.8.
The preceding computations were repeated for the example of Table
4.13 and Figure 6.8 with expected interval equal to 0.5 instead of 0.1. The
results are shown in Table 7.9. RASC distances ( X ) near the top and
bottom of the scaled optimum sequence now are based on fewer data (N*)
than those in the middle. In general, it does not make sense t o let the
jackknife estimator of position of an event be affected by events that are
clearly above or below this event. For this reason, a window equal to X f 2
was applied t o each cumulative RASC distance ( = X ) and events outside
this window were not used t o compute Q and s(Q). The reduced number of
pseudovalues ( = N ) used is also shown in Table 7.9. The width of the
window is such that N is approximately equal to N*. Setting the width

TABLE 7.8

Comparison of differences between successive values for example of Table 7.7. D and s(D)are intervals
and their standard deviations estimated by weighted scaling in RASC computer program. D1 and s(D1)
are corresponding jackknife estimates.

4-3 0 117 0 066 0 174 0 060

3-1 0 055 0 060 0 021 0 058
1-5 0 013 0 056 0 046 0 054
5-6 0 019 0 072 0 012 0 076
6-2 0 102 0 055 0 138 0 049

2-7 0 069 0 045 0 096 0 044

7-8 0 078 0 040 0 096 0 040

8~9 0 181 0 049 0 157 0 047

9-11 0 029 0 064 0 033 0 051
11-16 0 063 0 072 0 071 0 075

16-14 0 011 0 059 0 070 0 054

14 10 0 Ofi6 0 050 0 074 0 053
10 12 0 048 0 048 0 047 0 049
12-15 0 072 0 039 0 045 0 035
15-13 0 063 0 053 0 026 0 054
13-17 0 167 0 071 0 169 0 075
17-18 0 016 0 048 0 000 0 050
18-19 0 095 0 055 0 095 0 054
19-20 0 334 0 069 0 348 0 064
TABLE 7.9

Jackknife method applied to computer simulation experiment of Table 4.13 and Fig. 6.8. The 20 events
in 25 sequences have expected values E ( X ) spaced at intervals which are 5 times wider than those used in
the previous example of Tables 7.6 to 7.8. X,E(X),Q and s(Q) as in Table 7.7. The weighted distance
results X and Q were based on N* and N differences between successive 2-values, respectively. t(Y) is
+
studentized deviation of Y = X-E /X) 0.559.

-
I 0 000 0 000 0 000 0 000 I 0 559 ***
3 0 492 7 0 707 0 5.10 0 063 8 0 343 5 439**

2 0 507 9 0 354 0 637 0 096 9 0 712 7 439**

4 0 708 9 I 061 0 197 0 107 9 0 206 1925

5 1190 10 1414 I247 0 095 11 0 334 3 502**

6 I 393 10 1768 I442 0 160 12 0 184 1149

7 1843 14 2 121 1951 0 131 14 0 280 2 145*

8 2 069 13 2 475 2 146 0 164 14 0 153 0 908

9 2 505 14 2 828 2 476 0 168 15 0 236 1399

10 2 871 13 3 182 2 953 0 148 13
13 0 247 1665

11 2 977 13 3 536 3 053 0 139 0 000 0 000

12 3 287 13 3 889 3 34" 0 158 14 044 277

13 3 696 11 4 243 3 706 0 134 14 0 012 0 090
14 3 753 13 4 596 3 805 0 130 14 284 2 20*
15 4 234 13 4 950 4 407 0 096 12 157 163
16 4 261 13 5 303 4 406 0 Ill 12 484 4 63**

17 5 104 I1 5 657 5 349 0 189 9 0 006 0 031

18 5 153 8 6 010 5 413 0 162 9

10 299 I84
19 5 567 8 6 364 5 804 0 140 1a I Of

20 6 265 6 6 718 fi 509 0 220 4 0 I06 0 481

equal to 2 is equivalent to excluding events that occur above or below the

deleted event with a probability greater than 95 percent. In micro-RASC
(see Chapter lo), the user can change the width from its default value ( = 2)
t o any other value.
Both X and Q are relatively poor estimates of E(X) at positions near
the top of the scaled optimum sequence. Because these poor estimates
affect the other estimates lower down in the scaled optimum sequence and
the choice of origin is arbitrary, it was decided to reset the origin to the
position of event 11 near the midpoint of the scaled optimum sequence.
Consequently, studentized residuals t ( Y )were estimated for Y = X - E ( X )
+ 0.559 (see Table 7.9). As in Table 7.6, the majority of the studentized
residuals are within the 95 percent confidence limits. By means of
 257

TABLE 7.10

Jackknife method applied to Hay example. X, Q and slQi are weighted scaling results for cumulative
RASC distance, its jackknife estimate and jackknife standard deviation, respectively.

9 0.000 0.000 0.000

10 0.317 0.435 0.049
8 0.493 -.064 0.302
6 1.263 1.064 0.642
4 1.529 1.929 0.657
7 1.686 1.930 0.638
5 1.843 2.170 0.677
1 2.038 2.347 0.684
2 2.156 2.470 0.693
3 2.162 2.469 0.668

asterisks it is shown that some values of s ( Q ) ,especially those near the top
of Table 7.9, are too small. Although this indicates that, locally, there are
statistically significant discrepancies between X and E(X), these
differences are rather small in relative terms. In Table 7.7 the maximum
difference between X and E(X) is 0.254 or about 16 percent of the total
range ( = 1.598) of the RASC scale. In Table 7.9, the maximum difference
is 0.897 or 13 percent of total range (=6.718). It may be concluded that, on
the whole, the jackknife method yields good estimates of the positions of
the events in the scaled optimum sequence provided that the initial
ranking was good.
Table 7.10 shows Q and SCQ) in comparison with X for the Hay
example. The six events in the lower part of the scaled optimum sequence
are not only subject to strong clustering but also have relatively large
standard deviations. Events 8 , 9 and 10 clearly are above the other events
with events 8 and 10 having relatively small standard deviations. Event 6
may be intermediate between the preceding two groups. Differences
between X and Q for the Hay example are larger than those in Table 7.7
and 7.9. More research would be needed t o determine which estimate ( X or
Q ) is better than the other. It is known that jackknife estimators in
parametric estimation frequently are superior because bias of order n-l
(i.e. inversly proportional to sample size) tends to be eliminated (see e.g.
258

Miller, 1974). On the other hand, this advantage may be offset by the
introduction of bias related t o lack of stochastical independence of the
pseudovalues.
 259

CHAPTER 8

NORMALITY TESTING AND THE MODIFIED RASC METHOD

8.1 Introduction
The normality test of the RASC computer program was briefly
described in Section 6.6. In this chapter, it will be explained in more
detail. The problem of estimating the autocorrelation of the second-order
differences used in this test will be discussed first. A simple method will
be introduced by which it is possible to determine statistically whether or
not anomalous events belong to the normal distribution of the second-
order differences. For comparison with results obtained by Guex and
Davaud (1984)for a reworked bed using the Unitary Associations method,
the normality test will be applied to Drobne’s (1977)alveolinids from
Yugoslavia. The RASC computer program with normality test also will be
applied to Palmer’s (1954)data for the fauna of the Riley Formation of the
Llano Uplift in central Texas. Earlier, Shaw (1964)had constructed a
composite standard from Palmer’s database which involved t h e
determination and elimination of what he considered to be anomalous
events. It will be seen that the majority of the events deleted by Shaw are
not anomalous when the normality test is applied and this difference in
conclusions will be discussed.
The modified RASC method will be presented using the Gradstein-
Thomas database for example. This procedure can be used to construct
conservative range charts. Various types of range charts constructed by
different methods will be compared with one another in the last two
sections of this chapter. The modified RASC method can be very useful for
defining marker events which have variances that are much smaller than
the variances of other events. Modified RASC also provides new
information on the shapes of the frequency distributions of stratigraphic
events.
260

8.2 Autocorrelation of the second-order differences

The normality test was developed for two reasons: (a)to determine
anomalous events which in a specific section occur much higher or lower
than (at their average locations) in a regional standard developed on the
basis of a number of sections in a region; and (b) to test the normality
assumption used to transform cross-over frequencies into 2-values during
scaling. The normality test contributes useful information with respect to
both these objectives.
In the first few versions of the RASC computer program (Agterberg
and Nel, 1982; Heller et al., 19831, the simplifying assumption was made
that the second-order differences for stratigraphic events observed in
specific sections would be approximately normally distributed with
standard deviation equal to 20, if the original events are normally
distributed along the RASC scale with standard deviation equal to u. It
was realized that this simple model yields results which were at best
approximately true. In the original applications which were mainly to
Cenozoic and Cretaceous foraminifera1 databases for the northwestern
Atlantic margin, the final histograms of the normality test showed
observed frequencies that were, on the average, equal t o the expected
frequencies indicating that this simple model could be used. Three sets of
frequencies for the original normality test are shown in Table 8.1.
Anomalous events would cause observed frequencies of the highest
and lowest class (0,and Olo) to be greater than the expected frequency Ei
(i = 1,2, ..., 10) which is equal for all classes of i. During 1982 and 1983
when the RASC program was applied to other databases, several of which
were listed in Appendix I of Gradstein et al. (19851, it turned out that the

TABLE8.1

Normality test output from the original RASC program: Comparison of the observed frequencies (Oi)of
second order difference-values in each of the ten classes i = 1.2, ..., 10, with the expected frequencies (E,)
which are constant for each of the ten classes.

Source Ei Ol O2 O3 O4 O5 '6 '7 '8 '9 '10

Agterberg and Nel 24.1 27 23 26 20 27 24 28 22 21 23

( 1 9 8 2 , Table 6 )

HelleK e t al. 21.5 30 20 21 15 22 22 18 23 13 31

( 1 9 8 3 , Table 6 )

Gradatein 39.8 50 36 32 41 43 31 39 42 38 46
( 1 9 8 4 , Table 3 )
 261
TABLE 8.2

Normality test output for ten computer simulation experiments. Observed frequencies 0,are compared
to the expected frequency (=go) for each of the ten classes i = 1.2, ..., 10. E(D) represents the expected
interval (or RASC distance) between event-positions along the RASC-scale in these experiments.

O r i g i n a l RASC O1 O2 O3 O4 OS O6 O7 O8 O9 O10

E ( D ) = 1.0, Set I 156 55 32 69 127 145 54 39 48 175

E ( D ) = 1.0, Set 2 162 69 44 82 88 140 64 28 60 163
E(D) = 0.5, Set I 119 98 77 52 78 117 55 80 104 120
E ( D ) = 0.5, Set 2 119 95 89 62 79 94 84 59 88 131
E (0 ) = 0.3, Set I 89 111 75 84 87 88 85 77 107 97
E(D) = 0.3, Set 2 102 114 89 80 72 80 69 78 102 114
E ( D ) = 0.2, Set 1 84 101 83 76 80 98 106 100 88 84
E(D) = 0.2, Set 2 62 118 107 97 89 75 81 87 91 93
E(D) = 0.1, Set 1 18 77 91 135 115 123 153 111 62 15
E ( D ) = 0.1, Set 2 10 76 106 129 139 134 112 103 75 16

original normality test provided poor results in some situations because

the frequencies of anomalous events were either much larger or much
smaller than expected. For example, too many anomalous events were
found in the database for Baumgartner’s (1984) Jurassic Tethyan
radiolarians, and too few i n the Sullivan-Bramlette database for
Paleogene Californian nannofossils (cf. Section 4.2). It became difficult or
even impossible in these situations to define anomalous events on the
basis of the normal distribution model originally assumed t o hold
approximately true for the second-order differences. It was decided to
assess the problem systematically by means of the computer simulation
experiments previously described in Chapter 6.
Table 8.2 shows observed frequencies obtained by a pre-1985 version
of the RASC program for 10 classes of 900 second-order differences created
in ten of the computer simulation experiments previously described in
Chapter 6 . The expected frequency is 90 for all 100 entries for observed
frequencies in Table 8.2. Clearly, the observed frequencies in the tails of
these distributions for the second-order differences are too large when
E(D)is greater than 0.5 and they are too small when E(D)is less than 0.2.
It is noted that the runs for E(D) = 1.0, have a single greater than
expected frequency near the center of their distributions. This
phenomenon is related to the use of pairs of 2-values arbitrarily set equal
t o qc ( = 2.326) and with zero difference between them (see Chapter 6).
This constitutes a minor problem which is not related to the problem at
hand and does not arise for smaller values of E(D)in the experiments.
262

The applications of Table 8.1 may be compared to the experiments on

artificial data sets, with E(D) between 0.2 and 0.5, for which the observed
frequencies, on the average, are equal to the expected frequency Ei ( = 90)
in Table 8.2.
The present, revised normality test in the RASC computer program
consists of fitting a doubly-truncated normal distribution to the second-
order differences belonging to the classes with observed frequencies 0, t o
0,. If present, anomalous events are most likely t o occur in the tails of an
observed frequency distribution. Values in the classes of frequencies O,,
0,, 0, and O,,therefore were not used for estimating a theoretical normal
distribution.
Each second-order difference value in the normality test is computed
as follows. First, the difference of two successive values is calculated. If an
event precedes the next event for a section in the SEQ file, their difference
is corrected by subtracting a small amount. This correction is made
because a gradual increase in distance from the origin is t o be expected for
successive events in each section. The small amount was set equal t o the
difference between the highest and lowest cumulative RASC distance
values in the observed sequence for a section divided by the total number
of times an event precedes the next event for this section in the SEQ file
without being coeval to it. No correction is made for pairs of coeval events.
Next, the successive difference of two resulting values is determined. This
procedure resembles the calculation of a second derivative with respect to
location for every event except those in the first or last positions of a n
observed sequence.
The second-order difference calculated in the RASC normality test is
minus the difference between twice the RASC distance of an event on the
one hand and the sum of the distance of its two neighboring events, on the
other. If successive differences could be regarded as realizations of
independent normal random variables with variances equal t o 2u2, the
variance of the second-order difference would amount t o 6u2. This can be
seen as follows. Suppose that three successive distance estimates X 1 , Xk
and X k + 1 were normally distributed with zero mean and variance u’; then
the second-order difference 42Xk - X k - 1 - X k + 1 ) would be normal with
variance of 6u2 because u 2 ( 2 X k ) = 402 and u 2 ( X k - 1 ) = u 2 (Xk+l)
= u 2.
However, the successive distance estimates have become autocorrelated
because of the various manipulations to which the data were subjected
during ranking and scaling. Suppose that the autocorrelation coefficient
 263

of successive d i s t a n c e s Xk a n d Xk+l is w r i t t e n a s p w i t h
p = Cov (Xk,Xk+ 1)/u2. The variance of the second-order difference
satisfies

(8.1)

It follows that

0: = 202(p2-4p+3)
(8.2)

C o v ( X k - l , X k + l )= p202
if

The procedure followed in the RASC program consists of ordering the

second-order differences from all sections from the smallest to the largest
value. The standard deviation of the central 60 percent of the ordered
values is estimated and assumed t o represent a truncated normal
distribution. The relationship between standard deviations of truncated
normal and normal distributions is given in statistical tables. Their ratio
amounts to 0.463 if 20 percent is truncated from each tail. Division by
0.463yields the estimate 6,. Not all second-order differences are used for
this estimation because if anomalous values are present, these are more
likely t o occur in the tails of the distribution. From u2 = 3, it follows that
p can be estimated from 6, by

p = 2-41+03+
(8.3)

In general (cf. Agterberg, 1974, p. 3021, it can be assumed t h a t n

autocorrelated values are equivalent to n' stochastically independent
values with

I
lln' = l / n + 2 p d ( l - p ) - l / ( l - p ) 2
I /n2
(8.4)

This allows us to estimate n' which is part of the output of the RASC
program. In the chi-squared test for goodness of fit, expected frequencies
Ei of stochastically independent data in pclasses are related to the
corresponding observed frequencies Oi by

1=1 (8.5)
264
if t w o parameters of the fitted distribution were estimated. For
autocorrelated data, the sum on the left-hand side of this equation may be
multiplied by n'ln in order to obtain a n approximate estimate of
chi-squared.
The 10 classes of the normality test in the RASC program (cf. Section
6.6) were constructed by dividing the expected ordered sequence of second-
order differences into 10 equal parts in order to obtain 10 equal expected
frequencies for comparison t o the corresponding observed frequencies. The
class limits are given by the 2-values of the relative frequencies 0.1, 0.2,
..., 0.9 multiplied by 6,. This procedure provides a convenient normality
test. The individual second-order differences (top part of normality test
output as shown in Table 6.16) were compared to the 95% and 99%
confidence intervals k 1.960 6, and k 2.576 6,, respectively.
The preceding method generally yields sets of observed frequencies Oi
(i = 2,3,...,9) which are equal t o one another (and to Ei)except for random
fluctuations. The frequencies (0, and Ole) in the tails of the distribution
may be too high when anomalous events occur in several of the sections.
Results of applying the revised normality test for nine databases are
shown in Table8.3 and for six computer simulation experiments in
Table8.4. Other statistics for most of these computer runs are given in
Tables 8.5 and 8.6.
The normal distribution model provides a good fit for 13 of the 15 tests
in Table 8.3 according to the approximate chi-squared test (see last column
of Table 8.3). The 95 and 99 percent confidence limits of j;2(7)which should
not be exceeded if the normality assumption holds true (with levels of
TABLE 8.3

Revised normality test output for the nine databases in Agterberg et al. (1985) using RASC program.
Table 4.9 is slightly improved version ofdatabase 1; Tables 4.13,4.14and 4.15 are same as databases 9A,
9B and 9C, respectively.

1. Gradstein-Thomas 50.3 70 42 38 49 55 52 49 45 46 57 5.93

2. Gradstein 21.1 20 21 30 13 23 17 29 18 18 22 7.55
3, Doeven 64.1 78 53 65 68 53 64 70 67 53 70 3.36
4. Baumgartner 149.6 127 175 142 143 158 155 149 140 176 131 l>.80
5. Blank 172.2 235 139 145 139 210 173 179 147 118 235 53.72
6A. Rubel. brachiopods 62.3 61 59 65 73 52 59 66 69 63 56 2.35
6B. Rubel, ostracods 36.8 43 37 21 36 41 46 33 30 39 42 5.07
6C. Rubel. thelodonts 35.9 39 37 39 29 37 40 36 32 27 42 2.31
6D. Rubel, combined 57.6 50 75 45 62 62 51 54 69 52 51 12.52
7. Sullivan 47.4 55 40 40 49 66 37 44 46 42 55 2.45
8A. Corliss, tops 1.8 1 1 3 1 2 2 2 1 3 2 2.94
8B. Corliss, bottoms 5.0 6 2 6 10 5 3 2 7 2 7 9.24
9A. Agterberg-Lew, E(D)-0.5 45.0 44 41 56 35 34 35 42 54 48 41 8.76
9B. Agterberg-Lew, E(D)-0.3 45.0 43 45 45 44 38 57 36 56 50 36 6.20
9C. Agterberg-Lew, E(D)-0.1 45.0 62 29 34 46 43 46 53 47 50 40 3.51
 265

TABLE 8.4

Normality test output for six computer simulation experiments. See text for further explanation.

A. Revised RASC 01 02 03 0, 05 06 07 o8 0, ol0 X2(7)

(Set 1 only)

E(0) = 0.5 70 117 93 58 86 132 66 107 95 76 52.6

E ( D ) = 0.3 81 91 90 90 94 90 93 94 96 81 1.9
E(D) = 0.2 84 102 82 78 78 98 106 100 87 85 6.3
E ( D ) = 0.1 85 88 100 79 84 86 103 107 97 71 3.2

E ( D ) = 0.0, Set 1 98 90 73 86 94 98 83 120 85 73 0.5

E ( D ) = 0.0, Set 2 86 81 98 90 86 94 76 108 106 75 0.2

TABLE 8.5

Some statistics for RASC results for 9 databases of Table 8.3. The equivalent number ( n ' ) of
stochastically independent values was derived from number of second-order differences (n),standard
deviation 82 of Gaussian curve fitted to second-order differences (large values were not used, see text),
and estimated autocorrelation coefficient (0).

Data Base kc No. of No. of n 02 P n'

Events Sections

1. Gradstein-Thomas 7 44 24 503 1.223 0.420 206

2. Grad s t e i n 5 31 20 211 1.471 0.222 135
3. Doeven 7 77 10 64 I 1. I 0 8 0.508 210
4. Baumgartner 13 86 43 1496 1.701 0.027 1419
5. Blank 15 80 81 1722 I .419 0.264 1003
6A. Rubel, brachiopods 8 54 20 632 1.234 0.412 260
6B. Rubel, ostracods 8 40 12 368 1. I92 0.444 142
6C. Rubel, thelodonts 8 34 20 359 1.188 0.447 137
6D. Rubel, combined 13 43 35 576 1.659 0.063 507
7. Sullivan 9 52 10 474 0.791 0.725 76
8A. Corliss, tops 3 9 6 18 I .68b 0.040 17
8B. Corliss, bottoms 4 15 6 50 1.516 0.184 35
9A. Agterberg-Lew, E(D)=0.5 25 20 25 450 1.512 0.187 309
9R. Agterberg-Lew, E(D)=0.3 25 20 25 450 1.388 0.289 248
9C. Agterberg-Lew, E(D)=O.l 25 20 25 450 0.881 0.668 90

TABLE8.6

Autocorrelation statistics for RASC runs of five computer simulation experiments. If the original values
along the RASC-scale were stochastically independent, the ratio $2 I o would be equal to 1. Note extreme
reduction from n to n' for E(D) = 0.0. The negative autocorrelation coefficients 01 apply to second-order
differences (see text).

0.5 900 1.698 0.98 0.030 848 -0.658

0.3 900 1.528 0.88 0.173 634 -0.621
0.2 900 1.408 0.87 0.273 514 -0.597
0.1 900 0.966 0.56 0.609 219 -0.532
0 .0 900 0.327 0.19 0.948 25 -0.501
266

significance equal to 5 and l p e r c e n t ) , amount t o 14.1 and 18.5,

respectively. Only ^x2(7)= 53.7 of database no.5 clearly exceeds both
confidence limits. According to Blank (1984, p. 65) a number of events in
this database were determined to be anomalous because of four main
reasons: (1)taxonomic problems with Mesozoic events, (2) short sections
that were artificially truncated a t coring gaps, (3) contamination due t o
reworking, and (4) provinciality because of the large latitudinal spread of
control sites. The chi-squared value for database no.4 exceeds the 95
percent confidence limit but is below the 99 percent confidence limit.
There is the possibility t h a t the tail frequencies 0, ( = 127) and
O,, ( = 131) are slightly too small (in comparison with Ei = 149.6).
The run for E(D) = 0.5 in Table8.4 gave ;i2(7)= 52.6 indicating
nonnormality. It is likely that the central frequency 0, ( = 132) is
significantly greater than its expected value (Ei = 90) for the same reason
that 0, was too high in the computer simulation experiment with
E(D) = 1.0 (see Table 8.2).
In part B of Table8.4, the values of j12(7) are equal to 0.5 and 0.2,
respectively. The 1 and 5 percent confidence limits of 22(7) amount to 0.6
and 1.6, respectively. This suggests a degree of fit which is too good t o be
true. The approximate chi-squared test is based on the assumption that n
autocorrelated values are equivalent to n' independent values (see before).
As shown in Table 8.6, this reduction becomes very large (from n = 900 to
n' = 25) when E(D) = 0. There are no definite trends in the two sets of Oi-
values in Table 8.6. It may therefore be assumed that the procedure used
for estimating the observed and expected frequencies remains valid when
E(D) approaches 0 but that the reduction from n to n' has become too large.
Finally, it is noted that the autocorrelation coefficient fi estimated
from 62/0 applies t o the successive distances Xk and not t o the second-order
differences (Xk-l-Xk)-(Xk-Xk + 1). Suppose t h a t the autocorrelation
coefficient of the second- order differences is called pt. Then,

It follows that
p3- 4p2+ 7 p - 4
P, =
2p2-8p +6
 267

if cov ( x ~ +x,)~ =, p’02 i = 1,2,3

The latter condition would imply that the X k satisfy a first-order

Markov process (Agterberg, 1974). The autocorrelation coefficient p1 of
the second-order differences is negative and ranges from -0.6667 for p = 0
to -0.5 in the limit for p +l. Its values in five computer simulation
experiments are shown in Table 8.6. It is noted that the estimation of the
autocorrelation coefficients p and p1 has no bearing on the calculation of
the observed and expected frequencies of the normality test. The theory of
autocorrelation only was used to provide an approximate chi-squared test
for comparing the observed and expected frequencies with one another.
D’Iorio (1988) has performed experiments on the effect of increasing
the threshold value qc (=largest 2-value corresponding to P = 1.00) on the
RASC scaled optimum sequence for an integrated databank of Cenozoic
foraminifers and dinoflagellates on the Labrador Shelf-Grand Banks. The
total length for the scaled optimum sequence ( =maximum cumulative
RASC distance) increased from 7.781 to 12.351 when qc was enlarged from
its default value 1.645 (for P=0.95) to 2.576 (for P=0.995). When all
RASC distances, after enlarging q,, were reduced in length by the ratio
(7.781/12.351=) 0.630, there was little change in the shape of the
dendrogram. D’Iorio concluded that the scaled optimum sequence is not
sensitive to changes in the choice of q,. The large increase in qc in the
preceding experiment not only had a n undesirable effect on the total
length of the scaled optimum sequence, it also resulted in a slight but
significant distortion of the shape of the normal distribution of the second-
order differences. The estimated value of 62 (cf. Eq. 8.2), which amounted
to 1.454 (with 6 = 0.236) for D’Iorio’s 860 second-order differences with
qc = 1.645, increased to the unrealistically large value of 62 = 2.413 for
q,=2.576.
The latter value is too large because there is no reason to expect that p
in Equation (8.2)is much less than zero when n is too large. Consequently,
the upper bound of 02 is approximately d3=1.732 which is less than
62 = 2.413. By using q,-values that are too large, both u and 02 become too
large and Equations (8.3) and (8.4) are no longer valid. As a result, the
corrected sum used in the chi-squared test (cf. Eq. 8.5) was overestimated.
On the other hand, the 95% and 99% confidence limits for second-order
268

differences (used t o indicate possibly anomalous events in the normality

test for individual sections) are not sensitive to the choice of qc.

8.3 Unitary Associations and RASC methods applied to Drobne’s

alveolinids
Guex (1981) has coded biostratigraphic information on alveolinids
collected by Drobne (1977) and applied the Unitary Associations method to
these data. Information on 15 species in 11 sections a s used by
Guex (1981) is shown in Figure 8.1 and Table 8.7. Figure 8.2 from Drobne
(1977, Figs. 54 and 55, pp. 88-89) shows the original stratigraphic data for
one of the sections (11, Dane near DivaEa), for example. Forbidden
structures (see Chapter 3) have to be identified and eliminated before an
interval graph with Unitary Associations can be constructed from the
observed co-occurrences. The computer program of Guex a n d
Davaud (1984) i n i t i a l l y detected a s t r o n g component i n t h e
biostratigraphical graph for the Drobne data thus providing useful
information on biostratigraphical inconsistencies. This strong component
involved fossils 1, 3, 4, 11 and 13. The frequencies of arcs of the strong
component belonging to cycles C, were tabulated by Guex and Davaud
(1984) and the s-ratio (see Section 3.5)was determined. The arc from4 to 3
which occurs only in Section I (Fatji hrib) has the highest s-ratio ( = 3.00).
Other tabulations in the output from Guex and Davaud’s(1984)
computer program indicated that an abnormally large proportion of the
inconsistencies is due to the occurrence of fossils 3 , 4 and 8 in this same
section. In the original plot for individual sections (Fig. 8.1) it can be seen
that species 3 occurs higher in Section I than in the other sections where it
was observed. Drobne (1977, p. 83) specifically stated that bed no. 5 in the
Fatji hrib section which contains fossils 3 and 8 was reworked. For this
reason, Guex and Davaud (1984) decided to delete fossil 3 from their level
no. 4 in Section 1and t o repeat the analysis. Final results for the modified
computer run (without species 3 in Section 1)are shown in Table 8.8. The
method followed to obtain the unitary associations in the resulting “range
chart” was as described in Section 3.5. The five U.A.’s of Table 8.8 which
resulted from the union of some I.U.A.’s correspond closely t o the original
definition of Oppel zones (cf. Section 2.2).
In order t o illustrate the normality test, I previously applied it t o
Drobne’s alveolinids as follows (cf. Gradstein et al., 1985, pp. 253-262).
 LPlSAMl 1 2 3 4 5 6 7 8 9 10 11 12 13 14151L] IPISAMI1 2 3 4 5 6 7 8 9 1011 12131415lLl

1
11 7/ 1 1 1 '1

I!:I ; I
211-----

111 14 1
1 l
1 1
1 1
1 1
1
~

1
1
1
1
1 1
1
1 1 1
( I ) A. moussoulensis (9) A . montanarii
( 2 ) A. aramaea (10) A. aragonensis
( 3 ) A. solida
I (11) A . dedolia
(4) A. globosa (12) A . subpyreneica
( 5 ) A . avellana (13) A. laxa
( 6 ) A . pisiformis (14) A . guidonis
( 7 ) A . pasticillata (15) A . decipiens
( 8 ) A . leupoldi
Fig. 8.1 Occurrence of 15 alveolinids (1 to 15)from Yugoslavia (data from Drobne, 1977) in 11 sections
(I to XI). SAM: Sample numbers originally used by Drobne. Successive maximal horizons are numbered
in the stratigraphically upward direction for each section (see last column). Section XI is an isolated
occurrence described on page 92 of Drobne (1977). See Table 8.7 for names of sections.

TABLE8.7

List of sections for Drobne's dataset (cf. Fig.8.1).

I. Fatjihrib VII. Kozina-Socerb

11. Dane near DivaEa VIII. Golei
111. Veliko GradiSEe IX. Zbevnica
IV. RitomeEe near Gradisre X. Dane-Istria
V. Podgorje XI. JelSane (isolated sample)
VI. Podgrad-HruSica
270

:
1 .
I? I --

Marble
. ..
rn %:%lndles Flysch

Kozlna beds

Fig. 8.2 Drobne's (1977) original stratigraphic data for Section 11 in Fig. 8.1 (Dane near Divata). Circled
crass indicates stratum typicurn of new species. Samples 7,16,20 and 23 are for maximal horizons (Guex
levels).

The information of Table 8.1 was converted into RASC input by replacing
each fossil number i ( = 1, 2, ...,15) by two numbers (2i-1) for highest
occurrences and 2i for lowest occurrences, respectively. RASC was run on
the resulting data set with kc = 4, mcl = 1 and mc2 = 2. Setting kc = 4
ensured that no events were eliminated as in the U.A. computer program.
However, it became immediately apparent that 7 of the 15 species were
observed in one bed only in the sections containing them. Because the
highest and lowest occurrences of these 7 species coincided everywhere, I
decided to maintain a single number for each of these species indicating
occurrence only. (The odd numbers for these taxa indicate coinciding
highest and lowest occurrences.) Probabilistic ranking was applied and
followed by the modified Hay method. Three cycles occurred and each of
these involved the species 3 and 4. Based on mc2 = 2,42 out of 253 pairs of
 271

TABLE8.8

Final Unitary Associations (U.A.) for Drobne's alveolinids a s derived by Guex and Davaud (1984); upper
part of table is range chart with ones for taxa belonging to a particular Unitary Association; lower part of
table shows in which sections the final U.A.'s were identified.

1 0 0 0 0 0 1 0 1 1 1 1 0 0 0 1
2 0 0 1 1 0 1 0 1 1 1 1 1 1 1 0
3 0 0 1 1 0 1 1 0 0 0 0 0 0 0 0
4 1 0 1 1 1 0 0 0 0 0 0 0 0 0 0
5 1 1 0 0 0 0 0 0 0 9 0 0 0 0 0
U.A. Sections:
1 2 3 4 5 6 7 8 9 1 0 1 1
1 0 1 1 1 0 1 0 1 0 0 1
2 1 1 1 1 0 1 0 0 1 1 0
3 0 1 1 0 0 1 1 1 0 0 0
4 0 1 1 0 0 0 1 0 0 0 0
5 1 0 1 0 0 0 1 1 0 0 0

Explanation of numbers used for taxa: (1) A . mowsoulensis; (2) A. arumueo; (3) A. so/id(~;(4) A. glohosa; ( 5 ) A.
auelluna; (6) A. pisiformis; (7) A . posticillato; (8) A . leupoldi; (9) A. monfunarii;(10) A . aragonensis; (11) A. dedolio;
(12) A . suhp.yreneica: (13) A. luxu; (14) A . guidonis; (15) A . deciprens.

matrix elements were zeroed for scaling. Weighted distance analysis was
applied. From the results of the normality test (see Table 8.9),it may be
concluded that species 3 (A. solida) occurs too high in Section I (because of
reworking). In Table 8.9, A. solida has event number 5 for its lowest
occurrence (LO) which coincides with its highest occurrence (see before).

TABLE8.9

RASC normality test output for Drobne's Fatji hrib section with reworked bed at top (events 15 and 5
respresenting highest occurrences of fossils 8 and 3, respectively); the second-order differences were
tested for statistical significance; events with two asterisks are out of place with a probability of 99%;
those with one asterisk with a probability of 95%.

Event name Event RASC Second-order

number distance difference
LO A . leupoldi 15 0.626
LO A . solidu -5 2.660 -4.390 * *
LO A . subpyreneicu 23 1.550 2.911 *
HI A . pustic'illota - 14 2.172 0.023
LO A. pastidlata - 13 2.816 -2.589 *
LO A . glrhosu -7 0.871 1.871
HI A . pisijormis 12 2.044 0.492
HI A . pisiformis -11 2.962 0.239
LO A . urumucu 3 4.366
272

1,2; -15 lossil numbers

4 +-
I:2 - -, 5 unrrery aSSOCieb0"S ' I
average Ho (LAD)
4'5 I

Fig, 8.3 Comparison of RASC results to Unitary Associations for Drobne's alveolinids. Fossils were
ordered according to increasing RASC distance of their highest occurrence (HOor LAD).

Its RASC distance ( = 2.660) is larger than those of its neighbors in this
section. This discrepancy was brought out by computation of the second-
order difference (=-4.390**) in Table 8.9. The two asterisks indicate that
the event is out of place with a probability of more than 99 percent.
Figure8.3 shows a comparison of the 5 Unitary Associations of
Table 8.8 with the scaled optimum sequence used for obtaining Table 8.9.
The highest occurrences of the 15 fossils were ordered in Figure 8.3
according to their RASC distances. Because average highest and lowest
occurrences are estimated by scaling, the distances between them on the
RASC scale are less than their true stratigraphic ranges. According to the
original scaling model, events in sections are normally distributed about
their average position with standard deviations equal t o u = 0.7071.
Consequently, the observed highest occurrence of a fossil in a section
would occur with a probability of 95 percent below its RASC value
 273

decreased by 1.645 x u = 1.16. This value provides a more reasonable

estimate of the true highest occurrence or last appearance datum (LAD)
than the original RASC value. Likewise 1.16 can be added t o the RASC
distance estimated for a lowest occurrence in order t o obtain a more
conservative estimate of this lowest occurrence or first appearance datum
(FAD) along the RASC scale. The resulting enlargements of the RASC
ranges are shown as dashed lines in Figure 8.3. According t o the
probabilistic range chart of Figure 8.3, fossil 14 probably co-occurred with
3 and probably not with 2. The dashed lines are based on the assumption
that all events satisfy a normal distribution with the same standard
deviation along the RASC scale. I pointed before (Gradstein et al., 1985, p.
255) that this assumption may not hold true in reality and care should be
taken in interpreting the ranges of Figure 8.3. For example, Guex
(personal information, 1984) had advised me that fossil 5 probably never
coexisted with 11 although their ranges overlap in Figure 8.3.
The U.A. numbers of the fossils are also shown in Figure 8.3 and
circled if a fossil belongs t o a single U.A. only. The order of the
overlapping U.A.’s is very similar to that of the sequence of RASC ranges
for the fossils. The only discrepancy is that fossil 15 which belongs to U.A.
3 occurs in fifth position in Figure 8.3 while the other fossils of U.A. 3 ( 6 , 7
and 13) occupy positions 1 0 , l l and 12, respectively.
The preceding comparison using Drobne’s alveolinids is interesting in
that similar results for ranking as well as stratigraphic “normality” were
obtained by means of two methods (U.A. and RASC) which are built upon
different premises. In the U.A. method, observed co-occurrences of fossils
are augmented by virtual occurrences partly to resolve inconsistencies
(forbidden structures) in order to obtain assemblage zones. In the RASC
model, the observed highest and lowest occurrences of fossils in sections
are considered to be realizations of random variables with fixed average
positions along a linear scale. The two methods have in common that each
provides a way of eliminating inconsistencies and filling in the gaps due t o
missing data. In the U.A. method, this is done by adopting rules based on
graph theory whereas in the RASC method the observed data are
considered to belong t o small samples derived from (infinitely large)
statistical populations of which the parameters (rankings, means and
standard deviations) can be estimated.
The “zones” resulting from the U.A. method are primarily based on
observed and inferred co-occurrences of fossil species while the “zones”
274

resulting from the RASC method are primarily based on estimated

proximity of stratigraphic events i n time. Nevertheless, the two
approaches can yield similar results for anomalous occurrences and
groupings for correlation as shown in this section. It is noted that Guex's
maximal horizon method (cf. Section 4.5) was used for coding the
biostratigraphic information which implies loss of information from the
sequence file. During the past three years, the Drobne data have been
further discussed and re-analyzed by Guex (1987) and Brower (1989).
Moreover, because of the development of the modified RASC method, it
has become possible to construct range zones which a r e more
representative of the observed superpositional relations t h a n t h e
95percent confidence interval ranges shown in Figure 8.3. For these

TABLE 8.10

Alphabetic DIC file for Palmer's database. Numbers are for highest occurrences. Subtraction of one
gives code numbers for corresponding lowest occurrences. For example, 99 LO Angulotretu triangularis
is lowest occurrence corresponding to first entry (= 100) listed.

100 HI ANGULOTRETA TRIANGULARIS 44 HI HOLCACEPHALUS " E R U S

98 HT ANGULOTRETA TRIANGULARIS DIGITALIS 56 HI KINGSTONIA PONTOTOCENSIS
102 HI APHELASPIS CONSTRICTA 22 HI KINSABIA VARIGATA
104 HI APHELASPIS LQNGIFRONS 8 HI KORMAGNOSTLIS SIMPLEX
88 HI APHEUSPIS SPINOSA 108 HI LABIOSTRIA CONVMIMARGINATA
82 HI APHELASPIS WALCOTTI 116 HI LABIOSTRIA PLATIFRONS
120 HI APSOTRFTA MPANSUS 122 HI LABIOSTRIA SIGMOIDALIS
20 HI APSOTRETA ORIFERA 60 HI LLANOASPIS MODESTA
50 HI ARCUOLIMBUS CONVMUS 78 HI LLANOASPIS PECULIARIS
6 HI BOLASPIDELLA BURNETENSIS 66 HI LLANOASPIS UNDULATA
4 HI BOLASPIDELLA WELLSVILLENSIS 74 HI LLANOASPIS UNDULATA GRANULATA
10 HI CEDARINA CORDILLERAE 58 HI LLANOASPIS VIUGINICA
14 HI CEDARINA EURYCHEILOS 68 HI MARYVILLIA CF. M. ARISTON
84 HI CHEIMCEPHALUS BREVILOBA 76 HI METEORASPIS CF. M. LOIS1
94 HI CHEILOCEPHALUS MIWUTUS 16 HI METEORASPIS CF. M. ROBUSTA
34 HI COOSELLA BELTENSIS 54 HI METEORASPIS METRA
62 HI COOSELLA CF. C. WIDNERENSIS 12 HI MODOCIA CF. M. CENTRALIS
30 HI COOSELL4 GRANULOSA 2 HI MODOCIA CF. M. OWEN1
70 HI COOSIA CF. C. ALBERTENSIS 26 HI NORWOODIA QUADRANGULARIS
28 HI COOSIA CONNATA 40 HI OPISTHOTRETA DEPRESSA
64 HI CREPICEPHALUS AUSTRALIS 72 HI PEMPHIGASPIS INEXPECTANS
80 HI CREPICEPHALUS CF. C. IOWENSIS 106 HI PSEUDAGNOSTUS COMMUNIS
86 HI CREPICEPHALUS? PERPLEXUS 110 HI PSEUDAGNOSTUS JOSEPHUS
90 HI DICTYONINA PERFORATA 48 HI PSEUDAGNOSTUS? NORDICUS
52 HI DIERACEPHALUS ASTER 92 HI RAASCHELLA ORNATA
114 HI DU!?DERBERGIA VARIAGRANLIL4 38 HI SPICULE A
124 HI DYSORISTUS LOCHMANAE 24 HI SPICULE B
118 HI DYTRDUCEPHALUS GRANULOSUS 46 HI SPICULE C
112 HI DY"ACEPHALUS LAEVIS 18 HI SYSPACHEILUS CF. S. CAMURUS
32 HI GENEVIEVELLA CF. G. SPINOSA 42 HI TRICREPICEPHALUS CORIA
96 HI GERAGNOSTUS CF. G. TUMIDOSUS 36 HI TRICREPICEPHALUS TEXANUS
 275
reasons, the Drobne example will be recoded and subjected to modified
RASC later in this chapter.

TABLE 8.11

SEQ file for 7 sections of Palmer’s database. The event code numbers are explained in Table 8.10

MORGAN CREEK
119 -120 -123 -124 84 -100 -108 -114 82 -105 -106 101 -102 -103 -104 -113 90 -99 -107 87
-88 -92 81 42 -68 -83 -85 -86 -89 -91 69 -70 -77 -78 -79 -80 24 -65 -66 -67
-73 -74 40 8 -54 60 -64 38 -56 59 -62 63 22 23 -61 34 -49 -50 -52 -55
30 -39 -43 -44 -51 -53 19 -20 -25 -26 -27 -28 -29 -31 -32 -33 -35 -36 -37 -41
13 -14 -15 -16 -17 -18 -21 7 -10 9 5 -6

WHITE CREEK
120 113 -114 -117 -118 -121 -122 119 100 107 -108 82 -115 -116 99 92 -98 89 -90 -91
-97 45 -46 -81 24 -40 42 -56 -65 -66 -67 -68 59 -60 54 8 -36 22 33 -34
-41 -47 -48 -53 -55 -57 -58 35 27 -28 -39 21 -7.3 7 -13 -14 4 2 -3 1

JAMES RIVER
117 -118 100 82 -108 90 -97 -98 -107 81 -89 -99 24 -47 -48 -56 -68 -70 40 42
-77 -78 55 -63 -64 -65 -66 -67 -69 -71 -72 60 -61 -62 23 -59 8 -22 -30 -34
-50 29 -33 -35 -36 -39 -41 -49 7 -15 -16 -17 -18 -19 -20 -21

LITTLE LLANO RIVER

82 113 -114 99 -100 90 46 -89 -92 -93 -94 -95 -96 45 -83 -84 -91 42 -68 -69
-70 -77 -78 -81 -85 -86 40 24 -65 -66 -67 -73 -74 53 -54 -63 -64 56 34 -39
-41 -55 7 -8 -21 -22 -23 -33 -47 -48 10 9 5 -6

LION M O W A I N
84 -114 -118 -119 -120 82 -100 -106 -108 -112 -117 102 -104 99 -101 -103 -105 -107 -111 -113
81 -83 -87 -88 -91 -92 42 -68 -69 -70 67 7 -8 -31 -32 -34 -47 -48 -49 -50
-53 -54 -55 -56 29 -30 -33 -35 -36 -39 -40 -41 -43 -44 -45 -46

PONTOTOC
82 -100 99 107 -108 -109 -110 45 -46 -91 -92 -97 -98 8 3 -84 87 -88 81 41 -42
-68 67 -70 -75 -76 64 39 -40 -63 -69 22 21 -33 -34 8 -10 7 11 -12 9
6 5 3 - 4

STREETER
91 82 99 -100 92 81 -89 -90 40 -41 -42 -47 -48 -67 -68 -69 -70 -77 -78 24
-61 -62 33 -34 -53 -54 22 -23 -39 16 -18 -21 15 -17 14 7 -8 -13 9 -10
276

8.4 Application of RASC and normality test to Palmer’s database

for the Riley Formation in central Texas
Shaw’s (1964) book contains detailed documentation including a 126-
page appendix on construction of a composite standard for the fauna
(mostly trilobites) of the Cambrian Riley Formation of Texas originally
described by Palmer (1955). Various authors including Edwards and
Beaver (1978), Hudson and Agterberg (19821, Edwards (1982) and Guex

0.8219 0,6662 0,5104 0.3547 0.1990 0.0433

0,8991 0,7440 0.5881 0.4326 0.7168 0.1211 -0.0346
.............................................................
..................100 0.228\ HI ANGULOTRETA TRIANGULARIS
I
I ........ 87 0.06in HI APHELASPIS WALCOTTI
1 1
........................................................... 108 0.8651 HI LABIOSTRIA CONVEXIMARGINATA
I
I ............... 101 n.1813 LO LABIOSTRIA CONVEXIMARGINATA
1 I
I ....................................... 99 0.5445 LO ANGWOTRETA TRIANGULARIS
I I
I I .............. 90 0.1640 HI DICTYONINA PERFORATA
I I I
I I ................ 92 0.1941 HI RAASCHELLA ORNATA
I I I
I T ....................... 91 0.3031 LO RAASCHELLA ORNATA
I I I
I 1 I ................... H9 0.2451 LO DICTYONINA PERFORATA
I I
I 81 0.110R 1.0 APHELASPIS we.Lcnm
I I
I I 68 0.1307 HI MARYVILLIA CF. M. ARRISTON
I I
I I 10 0.0244 HI CWSIA CF. C. ALBERTENSIS
I I
I I 42 0.2641 HI TRICREPICEPHALUS CORIA
I I I
I 1 I . -.--
-. 69 0.0120 LO CWSIA CF. C. ALBERTENSIS
1 I I I
I I I .............. 24 0.1512 HI SPICULE B
I I I 1
I I I ............... 61 0.1186 LO MARYVILLIA CF. M. ARISTON
I I I 1
I I 1 1 ......... 40 0.0825 HI OPISTHOTRETA DEPRESSA
I I 1 1 1
I I 56 0.2128 HI KINGSTONIA PONTOTOCENSIS
I I I] ..............
I I 54 n.1652 HI METEORASPIS METRA
I
I
I
1
I
I
1
1 ____ 47 0.0000 LO PSNDAGNOSTUS? NORDICUS
I I I 1 I
I I T.. .................. 48 0,2517 HI PSEUDAGNOSTUS? NORDICUS
I I I1
I I I1 .... 55 0.0092 LO KINGSTONIA PONTOTOCENSIS
I I I1 I
I I 11 ............. 53 0.1429 LO HETEORASPIS METRA
I I I1 I
I I 11 I ...... 4 1 0.0281 LO TRICREPICEPHALUS CORIA
I I I1 1 I
I I 11 1 ........... 34 0.1051 H I COOSELLA BELTENSIS
I I I1 I I
I I 11 I I ........ 22 0.0684 HI KINSABIA VARIGATA
I I I1 I 1 I
I I ...................... 8 0.2864 HI KORMAGNOSTUS SIMPLEX
I
I
I
I
I
I _____ 23 0.0234 u) SPICULE B
I I 1 I
I I I ............. 39 0.1372 LO OPISTHOTRETA DEPRESSA
I I I I
I I ...................................... 33 0.5292 W CCOSELIA BELTENSIS
I I I
I I I ------- 21 0.0556 LO KINSABIA VARIGATA
I I

Fig. 8.4 Scaled optimum sequence (RASC 5/1/3run) for Palmer’s database for the Riley Formation in
central Texas.
 277

(1987) have used this database t o compare results obtained by .other

methods with one another and to Shaw’s composite standard.
Tables 8.10 and 8.11 contain DIC and SEQ files constructed from
Shaw’s Table A-1 (Shaw, 1964, pp. 230-232). Table 8.10 is an alphabetic
listing of highest occurrences of all fossils. The corresponding dictionary
numbers of the lowest occurrences are one unit less. Table 8.11 was
obtained after pre-processing of a DAT file (not shown here) with input
format as in Shaw’s table, and retaining only those events that occur in
five or more of the seven sections. Figure 8.4 shows the scaled optimum
sequence obtained after final reordering in a RASC 5/1/3 run. Input to
scaling was the optimum sequence resulting from probabilistic ranking.
(Although the modified Hay method also was applied, this did not affect
the probabilistic ranking results).
Table 8.12 gives the values of Kendall’s tau for the 7 sections in
comparison with the scaled optimum sequence. The seven tau-values
range from 0.74 t o 0.86 suggesting that all sections are correlated to the
average ranking with nearly the same strength.

Table 8.13 shows results of the overall normality test applied to the
180 second-order differences for events occurring in 5 , 6 or 7 sections. The
sum of the values in the last column is 3.163. This chi-squared value is not
statistically significant indicating that if there are anomalous events in
the sections, these are rare. Table 8.14 shows RASC normality test output
for the Morgan Creek, White Creek and Pontotoc sections.

TABLE 8.12

Kendall’s rank correlation coefficients for sequences of 7 sections correlated with scaled optimum
sequence of Fig. 8.4.

Section Tau
Morgan Creek 0.86
White Creek 0.81
James River 0.79
Little Llano River 0.80
Lion Mountain 0.74
Pontotoc 0.82
Streeter 0.75
278

TABLE 8.13

Overall normality test applied to Palmer’s database using taxa that occur in a t least 5 of the 7 sections.
No significant departures from normality are indicated.

ClassNo. 0 E 0-E (O-EWE

1 14 18 -4 0.415
2 19 18 1 0.026
3 26 18 8 1.659

4 18 18 0 0.000
5 16 18 -2 0.104

6 16 18 -2 0.104
7 17 18 -1 0.026

8 22 18 4 0.415

9 14 18 -4 0.415
10 18 18 0 0.000

TABLE 8.14

RASC normality test output for 3 sections in Palmer’s database. Only the lowest occurrences of
Tricrepicephalus coria and Opisthotreta depressa would be “too high” in the Pontotoc section. (Note that
both fossils occur in single beds in this section). Within the context of the entire database, these events
are not anomalous because, on the average, 4 single star events and 1 double star event are expected to
occur in every set of 100 events.

MORGAN CRFEK CUM. DIST. 2ND ORDER DIFF.

H I ANGLILOTRETA TRIANGULARIS 100 0.0000

H I LABIOSTRIA CONVMIMARGINATA -108 0.2955 -0.1411
HI APHELASPIS WALCOTTI 82 0.2285 1.1249
HI DICTYONINA PERMRATA 90 1.8865 -1.8172
LO ANGLILOTRFIA TRIANGULARIS -99 1.3ldO 0.3631
LO LABIOSTRIA CONVEKIMARGINATA -101 1.1606 0.6859
HI RAASCHELLA ORNATA 92 2.0504 -0.1476
LO APKELASPIS WALCOTTI 81 2.1926 0.1837
HI TRICREPICEPHALUS CORJA 42 3.7185 -0.6951
HI MARYVILLIA CF. M. ARISTON - 68 3.5635 -0.8609
LO DICTYONINA PERFORATA - 89 2.5416 0.1128
LO RAASCHELLA ORNATA -91 2.2445 1.6560
LO CWSIA CF. C. ALBERTENSIS 69 3.9826 -1. 6414
HI CWSIA CF. C . ALBERTENSIS -70 3.6942 0.2637
HI SPICULE B 24 4.0546 0.1820
LO MARYVILLIA CF. M. ARISTON -67 4.2118 -0.3638
HI OPISTHOTRETA DEPRESSA 40 4.1905 0.9479
H I KORMAGNOSTID SIHPLEK 8 5.5110 -1.5125
HI MFTEORASPIS HETRA - 54 4.1451 0.1132
HI KINGSTONIA PONTOTOCENSIS 56 4.4130 1.2483
H I KINSABIA VARIGATA 22 5.4485 -0.6207
LO SPICULE B 23 5.8034 -0.8154
HI COOSELL4 BELTENSIS 34 5.3429 0.6655
LO KINGSTONIA PONTOTOCENSIS -55 5.1626 0.4592
LO OPISTHOTRXTA DEPRESSA 39 5.8268 -0.9340
LO METEORASPIS -53 5.1719 1.0620
LO COOSELLA BELTENSIS 33 5.9641 -1.0563
LO TRICREPICEPHALUS CORIA -41 5.3148 1.4425
LO KINSABIA VARIGATA 21 6.4933 -1.1228
LO KOKMAGNOSTUS SIHPLM 1 6,5489
 279

TABLE 8.14(continued)

WHITE CREEK CUM. UIST. 2ND ORDER 1IIFF.

HI ANGULOTRETA TRIANGULARIS 100 0.0000

LO LABIOSTRIA CONVMIMARGINATA 107 1.1606 -1.5892
HI LABIOSTRIA CONVMIMARGINATA -108 0.2955 0.3616
HI APHELASPIS WALCOTI'I 82 0.2285 1.1804
LO ANGULOTRFIA TRIANGULARIS 99 1.3420 -0.4050
HI RAASCHFLLA ORNATA 92 2.0504 -0.2113
LO DICTYONINA PERFORATA 89 2.5476 -0.7717
HI DICTYONINA PERFORATA -90 1.8865 1.0191
LO RAAScHnLA ORNAlA -91 2.2445 -0.2465
LO APHELASPIS WALCOTTI 81 2.7926 0.7138
HI SPICULE B 24 4.0546 -0,4895
HI OPISTHOTRETA DEPRESSA -40 4.3905 - 1.4444
HI TRICPJZPICEPHALIIS CORIA 42 3.7185 1.8630
HI KINGSTONIA WNTOTOCENSIS -56 4.4730 -1.0156
LO MARYVILLIA CF. M. ARISTON -67 4.2118 -0.3872
HI MARYVILLIA CF. M. ARISTON -68 3.5635 1.394c
HI METEORASPIS METRA 54 4.7451 -0.4111
HI KORMAGNOSTUS SIHPLM 8 5.5170 -0.8396
HI KINSABIA VARIGATA 22 5.4485 0.5840
LO CCOSELLA BELTENSIS 33 5.9641 -0.1001
HI COOSELLA BELTENSIS - 34 5.3429 O.5Y31
LO TRICREPICEPHALUS CORIA -41 5.3148 -0.3758
LO PSEUDAGNOSTUS? NORDICUS -47 4.9109 0.4038
HI PSEUDAGNOSTUS? NORDICUS -48 4.9109 0.2609
LO METEORASPIS m RA -53 5.1719 -0.2 701
LO KINGSTONIA PONTOTOCENSIS -55 5.1626 0.2368
LO OPISTHOTRETA DEPRESSA 39 5.8268 0.0022
LO KINSABIA VARIGATA 21 6.4933 -0.9197
LO SPICULE B -23 5.8034 0.9988
LO KORMAGNOSTUS SIMPLM 7 6.5489

PONTOTOC CUM. DIST. 2NU ORDER DIFF.

HI A P W S P I S WALCOTTI nz 0.2285
HI ANGULOTRETA TRIANGULARIS -100 0,0000 0.9959
LO ANGULOTRBXA TRIANGULARIS 99 1.3420 -1.5233
LO LABIOSTRIA CONVMIMARGINATA 107 1.1606 -0.1092
HI LABIOSTRIA CONVEXIMARGINATA -108 0.2955 2.2396
LO RAASCHELLA ORNATA 91 2.2445 -1.5685
HI RAASQiELLA ORNATA -92 2.0504 0.3617
LO APHELASPIS WALCOTTI 81 2.7926 1.7199
LO TRICREPICEPHALUS CORIA 41 5.3148 -3.5439 W
HI TRICREPICEPUALUS CORIA -42 3.7185 1.4412
HI MARYVILLIA CF. M. ARISTON - 68 3.5635 0.2288
LO MARYVILLIA CF. M. ARISTON 67 4.2118 -0.5914
HI COOSIA CF. C. ALBERTENSIS -70 3,6942 2.0758
LO OPISmOTRETA DEPRESSA 39 5.8268 -2.9945 91
HI OPISTHOTRBXA DEPRESSA -40 4.3905 1.0286
Lo CWSIA CF. C. ALBERTENSIS - 69 3.9826 1.2991
HI KINSABIA VARIGATA 22 5.4485 -0.4212
LO KINSABIA VARIGATA 21 6.4933 -0.9993
Lo CO0SET.l.A BELTENSIS -33 5.9641 -0.0920
HI CDOSELLA BELTENSIS -34 5.3429 0.2207
HI KORMAGNOSTUS SIMPLM 8 5.5170 0.8579
LO KORMAGNOSTUS SIMPLM 7 6.5489

To those who have read Shaw's (1964) book, the preceding evaluation
of Palmer's database may seem surprising in that during his construction
of the composite standard, Shaw frequently did not use events which were
deviating more than other events from the straight lines fitted by the
280

method of least squares to events initially in two sections plotted against

one another, and later in other sections plotted against the composite of
two or more sections. However, most of these unused events appear not t o
be anomalous in a statistical sense. It may be concluded that Shaw was
trimming the data in order to improve least-squares estimation of the lines
of correlation. Trimming is a statistical procedure in which estimates are
restricted to measurements which are relatively close to the quantity to be
estimated. Such methods now are widely used in exploratory data
analysis (Tukey, 1977). It is noted that, in order to obtain the normal
distribution of the second-order differences, only 60 percent of the
observations were used (see Section 8.2). This can be regarded as another
example of trimming.

It will be shown in Section 8.9 that Shaw’s composite standard

method, because of trimming, yields a range chart with ranges that, for
some taxa in length are intermediate between those in the scaled optimum
sequence of Figure 8.4 and extended ranges resulting from the modified
RASC method with use of all observations. On the whole, however, the
ranges obtained by modified RASC are very similar t o those obtained by
other “conservative” range chart construction methods including the
composite standard method.

8.5 Modified RASC Method

Although robustness is increased by combining events with one
another (application of central limit theorem, see Chapter 61, ordinary
scaling is based on the assumption t h a t all events have normal
distributions with equal variance along the interval scale. It is noted that
the assumption of equality of variance for different events frequently has
been made in quantitative stratigraphy in a n implicit manner. For
example, Shaw’s (1964) lines of correlation were fitted assuming that this
condition is satisfied.

By comparing individual sequences with the scaled optimum

sequence and collecting deviations from smoothing splines fitted for
different sections, it is possible to estimate the frequency distribution of
each event separately. The RASC scaling algorithm can be modified to
allow for different variances of the events. An iterative procedure has
been developed (cf. Agterberg and D’Iorio, in press; D’Iorio, 1988; D’Iorio
and Agterberg, 1989) in which the methods of (1) weighted spline fitting,
 28 1

and (2) modified scaling are applied alternately until a stable solution is
reached upon convergence. In these two methods, the variances of the
events are not assumed t o be equal to one another. Application of this
method t o highest occurrences of Cenozoic foraminifers along the
northwestern Atlantic Margin (Gradstein-Thomas database) showed
(1) unequality of variances for different events; and (2) minor departures
from normality of the frequency distributions for separate events.
Changes in the scaled optimum sequence resulting from the iterative
procedure were negligibly small. The new approach allows identification
of small-variance e v e n t s which d i s a p p e a r e d a p p r o x i m a t e l y
simultaneously from different sections in the same study region.
The RASC method for ranking and scaling consists of (1) forming a
single, optimum sequence from mutually inconsistent sequences of
observed events for different stratigraphic sections, and (2) positioning
these events along a relative time interval scale. In modified RASC, the
scaling part of the RASC method is generalized t o account for possible
differences in uncertainty associated with the positioning of different
events along the RASC interval scale. The original scaling model was
illustrated in Figure 6.4. Each of a group of biostratigraphic events (A, B,
..., G) was assumed to be a random variable (XA,XB, ...,XG)with Gaussian
probability distribution along the RASC scale. These Gaussian curves
have different means (EXA, EXB, ..., EXG) but their variances (u2) are
assumed to be equal to one another. By means of this model it became
possible to estimate the intervals between the successive mean values
denoted as EXA, EXB, ...,EXG. The model of Figure 6.4 can be generalized
by allowing the variances of the events t o be different. Such an extension
of the method only is possible if the variances CJA,UB, ..., OG of the
frequency distributions ~ ( x A )flxg),
, ...,~ ( x G of
) the events can be estimated.
A possible estimation procedure is described here.
The original RASC method provides estimates xi of EXi where i
denotes events. In each stratigraphic section xi can be plotted against ui,
representing relative position of event i in the so-called event level scale of
the section. New estimates fi of EXi in the section can be obtained by
fitting a cubic spline curve with u as the independent variable. The
differences (+xi) can be collected from all sections in which event i occurs
and plotted as a histogram that provides an approximation of flxi-EXi).
The shape of the latter distribution is the same as that of f l x i ) . The
standard deviation Si of the differences provides an estimate of oi.
282

In the application to Cenozoic Foraminifera from 24 wells on the

Labrador Shelf and Grand Banks t o be discussed in the next two sections,
distinct differences were found i n the widths of the probability
distributions f l x i ) for different events. The number of differences per event
(sample size, n) varies from 7 to 22 in this application. Most observed
frequency distributions are unimodal and slightly skewed to the right or t o
the left. A few distributions may be bimodal. The sample sizes are too
small t o demonstrate statistical significance of the possible departures
from the Gaussian model. However, each event can be assumed to have its
own variance because the widths of the f l x i ) are clearly different. This led
to the modified RASC model to be explained in this section. Application of
modified RASC with different variances for different events, results in a
new set of estimates of the positions of EXA, EXB, ..., EXG. Spline-curves
can again be fitted to data for individual sections. Repetition of these steps
results in an iterative procedure which converges toward a final solution.
The histograms of the differences (12i-q) after convergence provide better
approximations of f ( x J than the histograms a t the beginning of the
iterative process.
Suppose that the x-axis for relative time interval scale points in the
stratigraphically upward direction. For example, the events A, B, ..., G in
reversed order, may represent highest occurrences encountered
successively in a well drilled downward in a basin where age increases
with depth. The location of each stratigraphic event is represented as a
random variable (XA,XB, ..., XG) that in each well may assume a specific
value along the x-axis with probabilities controlled by its Gaussian curve.
Suppose that two events (e.g. A and B) both occur in R wells. In R A wells
A is observed above B and in R B wells B above A. When A and B are
observed t o be coeval in a well, 0.5 is added t o R A as well as t o RB. Setting
+
R A RB = R , the ratio PAB = RA/R can be set equal to the probability
that A is observed before B in a randomly selected well and used to
estimate the interval AAB = EXB-EXA. The difference AAB is the mean of
a random variable DAB = XB-XA for difference between the random
variables X B and XA. If AAB is positive, DAB would turn out t o be positive
in most sections. However, the model also allows B to be observed before A
in some sections with negative DAB. If the Gaussian curves of two events
were t o coincide, the probability that one of these two events is observed
before the other, is exactly 0.5. If the variances of the Gaussian curves in
Figure 6.4 are all equal t o a2, PABestimates
 283

(8.8)

In this equation, which is equivalent to Equation (6.1), the mean interval

AAB is divided by a d 2 representing the standard deviation of the random
variable DAB. If the RASC model, it is not possible to estimate both AAB
and u. For t h i s reason, CJ was set equal t o a n a r b i t r a r y constant
(u = 0.7071). A different choice of u would be equivalent to rescaling the
axis for the distance estimates (x-axis). From Equation (8.8)it follows that
AAB = @-' (P(DAB>O)}. Consequently, ZAB = @-~(PAB) where PAB is
converted into ZAB representing a fractile of the normal distribution in
standard form. Suppose now that events A and B have different variances
U ~ B . Then the variance of DAB becomes u AB = u A + u B. The
U ~ and
A
2 2 2
corresponding standard derivation UAB reduces to 0 4 2 = 1 only if
CJ~A = U ~ B=02. In the modified RASC model, Equation (8.8) is replaced
by

and ZAB is replaced by GAB = ZAB-SAB. Thus, t h e ZAB-value of a

relative frequency PAB must be multiplied by SAB representing a n
estimate of UAB before it can be interpreted as a n estimate of the interval
EXB-EXA. As pointed out before, the precision of a Z-value depends on
relative frequency P as well as sample size R . More weight w can be given
to G-values with larger R by using the equation

(8.10)

where s2(G)denotes estimated variance of G. These weights may be used

when sets of G-values are combined with one another in order to improve
the estimate of the interval between two events. For example, because
(EXc-EXA) - (EXC-EXB) reduces to EXB-EXA, GAB.C = GAC - GBC
provides a n indirect estimate of EXB-EXA w i t h weight W A B . C =
(WACXU.JBC)/(WAC + wgc). The direct estimate GAB can be combined
with GAB.C and other differences between G-values according to the
equations (e.g. Eq. 6.2) previously used for the Z-values.
284

8.6 Application of modified RASC to the Gradstein - Thomas

database
The database used in this example is for highest occurrences of
Cenozoic Foraminifera in 24 exploration wells on the Labrador Shelf and
Grand Banks previously introduced in Section 4.6 (see Tables 4.7 and 4.9).
Table 8.15 shows estimated RASC distances for 44 events each occurring
in at least 7 wells. This RASC distance is plotted against event level in
Figure 8.5A for one of the wells (Adolphus D-50). The horizontal scale for
relative event levels increases with depth. The Adolphus D-50 well was
sampled by taking cuttings a t a regular interval of 30 ft (approximately
10 m). Only 23 distinct levels t o a depth of about 9000 ft showed one or
more highest occurrences for the 44 species considered. These levels were
numbered from 1 t o 23 in Figure 8.5. In total, only 30 of the 44 species
were encountered in Adolphus D-50.
A cubic spline curve was fitted to the data shown in Figure 8.5A with
smoothing factor set equal to u = 0.7071 representing the standard
deviation of events along the distance scale in the ordinary RASC model
(see before). In general, the smoothing factor (SF) is the square root of the
mean squared deviation for the deviations between points and spline curve
(measured along the RASC distance scale). SF is selected in advance and
the best-fitting spline curve will have SF as standard deviation (biased
estimate) of its residuals. This standard deviation is “biased” because the
sum of squares of the deviations was divided by n instead of its number of
degrees of freedom. For example, the number of degrees of freedom for a
best-fitting straight line is n-2. Division of the sum of squared deviations
by n-2 then results in an “unbiased” estimate. The best-fitting straight
line is the smoothest possible spline-curve. This solution always is
obtained if SF exceeds the standard deviation of the residuals from the
best-fitting straight line. If the spline-curve is not a straight line, the
number of degrees of freedom is not readily determined. An unbiased
estimate of SF could be obtained by cross-validation (see Section 9.5) but
this method is not used here.
In the original RASC model, it is assumed that all events have the
same standard deviation (0).In modified RASC, each event i has its own
standard deviation ui estimated from the n deviations of the event in the
wells where it occurs. The sum of squared deviations for each event was
divided by (n-1) to obtain the estimated variance si2 (see Table8.15,
3rd column). This is an “unbiased” estimate because, in general, the
TABLE 8.15

RASC distances and variances si2 estimated for 44 species (event numbers as in Gradstein et al., 1985)
before (First run) and after (Fifth and Sixth runs with refinement) convergence.

I.'IRBI'KUU

Unhi.rsed I.'ohiascd L'nhiascd

Event IMSC Event HASC I:"C,ll IMSC
"url.llleC "ill ,*rice Yalld"Cc
nuinher dlil nurnhcr tlibl "lllllbVI dist
(0 mean) 10 mean) 10 lllennl
~ ~ ~ ~ __ ~

10 0 000 11 978 I0 11 ono I I167 I0 n OIIO I057

17 o 288 0 688 17 0 4.1I 0 699 17 0 439 0 702

16 I016 0 341 16 1137 0 266 I6 1138 0 2RI

67 I237 0 511 67 1216 0 557 67 1215 0 524

18 1616 0 202 18 I669 0 1195 I8 I665 I1 093
21 I858 0 085 21 I722 0 016 21 I715 0 009
71 1865 0 427 20 I 837 0 073 20 I 830 0 070

20 I946 0 164 71 I855 11 310 71 I 818 0 372

26 2 087 0 3% 26 I 983 0 409 26 I97G 0411
70 2 337 0 145 70 2 171 0 121 70 2 167 0 135
15 2 370 0 446 15 2 206 0 412 15 2 199 0419

24 2 754 0 199 24 2 573 0 173 24 2 567 0 180

27 2 768 0 649 27 2 724 I1 725 27 2 720 0 735
69 2 988 0 649 69 2 869 0 636 69 2 862 0 632
25 3 084 0 319 25 2 894 0 23s 25 2 890 0 238

81 3 168 0 5B2 81 3 007 0 615 81 3 000 I1 624

202 3 289 o 28s 2112 3 144 0 110 20 2 3 141 0 1193
259 3 400 11 151 259 3 236 0 092 259 3 233 n 092
34 3 834 n 4.19 I47 3 668 0 173 147 3 667 0 166

147 3 898 0 413 34 3 718 0 537 34 3 717 I) 554

33 4 om Inm 33 3 833 1111 33 3 861 I142

260 4 I14 0 1911 260 4 007 I1 149 260 4 0117 n 151

261 4 I55 0 134 261 4 133 0 068 26 I 1 1.14 0 070

263 4 297 I1 347 263 4 187 0 339 26.1 4 I88 n 350

29 .I 520 0 "12 29 4 3z2 n 136 29 1321 n I.IS

32 4 603 0 2n9 32 4 419 0 218 12 1420 I1 '?20

4n 4 662 I1 554 40 4 441) I1 426 Ill 4 -437 n .133

261 4 869 0 161 .12 4 682 0 824 42 4 680 o a43
42 -I an2 0 7?9 264 4 691 11 355 21i4 4691 I1 :159

311 4 921 n .$Fin .I I 4 735 I1 352 41 4 735 I1 361

11 i947 I1 496 111 4 798 I1 I99 311 4 799 II4lfi

90 5 235 0 368 90 5 041 0 384 911 5 1143 0 413

86 5 249 0 175 86 5 053 I1 1142 36 5 052 0 377
36 5 315 0 332 36 5 056 0 356 86 5 053 n 033
57 5 352 11 son 57 5 1195 0 544 57 5 095 0 557

.15 6 906 0 819 45 5 655 0 916 45 5 653 u 92s

50 6 Ill1 11 2114 50 5 886 0 no8 50 5 885 11 10112
46 6 227 U 597 46 5 926 11 397 46 5 923 0 393
230 6 :125 0 132 230 6 053 11 :197 230 li 051 0 395
52 6 426 I1 5511 54 R nii7 0 217 54 ii 1167 0 222

54 6 473 I1 Z(i7 52 6 I30 1) 174 52 6 iza 11 ilia

56 6 925 I1 3;2 56 6 1Xfi 0 I95 sii 6 385 I1 I89
I1 274 0 261 rr 9.10 I1 2 i f i
55 7 405 65 6 Y37 .A> fi

59 7 780 0 57G 59 7 I(i4 11517 5') i 162 11 515

286

-I I 3 5 1 9 I1 13 15 11 19 21 23 25 -I I 3 5 1 9 I1 13 15 11 19 21 23 25
Level LQVQ~

Fig. 8.5 Results of fitting a spline-curve to data for Adolphus D-50well before (A) and after (B) iteration.
For Fig. 8.5A, the smoothing factor (SF) was set equal to SF=0.7071 and standard deviations for
individual data (si) were kept equal to 1.000, This procedure provides results identical to setting
SF= 1.000 and s,=0.7071 for all i). For Fig. 8.5H,the smoothing factor was set equal to S F = 1.000 and
use was made of s,-values obtained after convergence. In both diagrams, SF exceeded the standard
deviation of the residuals so that the spline-curve became a best-fitting stratight line.

number of degrees of freedom for n deviations from a mean is equal to n-1.

The values of si2 could be used to run the modified RASC program. This
would give a different set of RASC distances which, in turn, might be used
to estimate new variances from new spline-curves. However, the values of
si2 also can be used to repeat the spline-curve fitting stage without first
going through modified RASC.
In weighted spline-curve fitting, the observations are weighted
according to the inverse of their variance. Application to Adolphus D-50
using the values of si2 in Table 8.15 (3rd column) yielded a n improved
best-fitting straight line. Deviations from this line and spline-curves for
the 23 other wells gave improved estimates si2 which were used as input
for modified RASC. This extra step is only taken at the beginning of the
iterative process. During later steps, weighted spline-curve fitting is used
only. It was found that the iterative process converged t o the same final
solution with and without the extra step a t its beginning. With this
refinement, the final solution was reached faster. Modified RASC
distances and the variances used to obtain them are shown in Table 8.15
for steps 5 and 6 of the iterative process with refinement. These estimates
are preceded by their fossil event numbers because of minor reordering
with regard t o the original sequence order (Table 8.15,column 1). The
weighted spline-curve fitted after step 5 of the iterative process with
refinement for Adolphus D-50 is shown in Figure 8.5B.
 287

At the beginning of the iterative process, the average variance for the
44 species is equal to 0.500. A t the end of the process the overall variance
has become 0.351. This implies that the standard deviation u = 0.70 was
reduced to 0.59. The total range for the species along the RASC scale was
reduced from 7.78 (original RASC output) to 7.16 after steps 5 and 6 (cf.
Table8.15). This shrinking is related to the reduction in the standard
deviation.
The mean deviation of the species in individual wells from their
spline-curves was computed a t each step of the iterative process. In
Figure8.6, this mean deviation is plotted against RASC distance at the
beginning (RASC output) and end of the iterative process (modified RASC
output). Clearly, there is a systematic departure from zero near the top
and bottom of the stratigraphic sequence. The average deviation of the
first 3 species amounts to -0.65 and that of the last 9 species is 0.28 in
Figure 8.6B. The discrepancies for these 12 events were not significantly
reduced during the iterative process. It indicates that, on the average, the
fitted spline-curves slightly underestimated RASC distances near the tops
of the sections and overestimated them near the bottoms. This effect
would be reduced if more weight were given to the 12 events, e.g. by
centering their variances with respect t o the average deviations.
However, this also would result in a further decrease of the overall
variance with increased shrinking of the total range for the species along
the RASC scale.

8.7 Frequency distributions of stratigraphic events

As mentioned in the previous section, most frequency distributions
for individual species are unimodal and slightly skewed to the right or t o
the left. A few distributions seem t o be bimodal. All distributions change
shape during the iterative process. We will restrict our presentation
mainly to the final result obtained after convergence.
Figure 8.7 shows histograms for taxon 42 (Cibicidoides alleni) and
taxon 50 (Subbotina patagonica) before and after convergence. S.
patagonica which is an abundant planktonic species w a s already a
relatively good marker at the beginning of the iterative process because its
variance ( = 0.204) was less than 0.5. After convergence, its variance has
become very small. The corresponding histogram is a narrow peak
indicating that the final spline-curves for the nine wells with S.
288

A I
i
V
c
P
a
8.5
9
e

d
I
f
f I 00

e
P
e
n -8.5
-
C
e Foraminifera of the Grand Banks
and Labrador shelf

e
r
a
9
8.5 1
i

d
I
f
f
e
r
e
n -8.5 .:
C
e Foraminifera of the Grand Banks
and Labrador shelf

Fig. 8.6 Mean deviation from spline-curves per species plotted against RASC distance before (A) and
after (B) convergence. For further explanation see text.

patagonica passed almost exactly through the points for this taxon. It may
be concluded that S.patagonica is an excellent marker, whose position in
individual sections is everywhere close t o its position in the scaled
optimum sequence. This property is enhanced when modified RASC is
used. On the other hand, Czbicidoides alleni which is a rare benthonic
species has a variance above 0.5, both before and after iteration. Its
histogram also has not changed significantly (see Fig. 8.7). This taxon
seems t o have a bimodal frequency distribution. According t o
F.M. Gradstein (personal communication, 1987), C. alleni is not well
defined taxonomically and may actually represent two different forms.
 289

An unsolved problem of considerable interest regards the shapes of

unimodal frequency distributions of biostratigraphic events. It is unlikely
that such frequency distributions are exactly symmetrical. Two models
with asymmetry for highest occurrences were suggested in Section 2.6:

Model A -The species disappeared in most places at approximately

the same time but, perhaps due to lack of preservation, had already
disappeared earlier i n some places. This is the most likely model for exits
as explained in Section 2.6. A “mass extinction” or a hiatus would create
frequency distributions of this type. Model A predicts negative skewness
(cf. Fig. 2.10D).

Model B - The species disappeared in most places (from most sections)

at approximately the same time but remained in existence longer in a few
places due to favorable conditions or was subjected to localized reworking.

Event nvlber 50 : SUBBOTINA PATAGONICA

-1.5 -1.1 -0.7 -0.3 0.1 0.5 0.9 1.3 -1.5 -1.1 -0.7 -0.3 O.! 0.5 0.9 1.3
-1.3 -0.9 -0.5 -0.1 0.3 0.7 1.1 1.5t -1.3 -0.9 -0.5 -0.1 0.3 0.7 1.1 1.5t
DIFFEREKE DlFNlwtE

Event n u h r 42 : CIBlClWIDES ALLEN1 Evmt n u h r 42 : CIBICIWIES NLENl

7
*

3 .. - 3 .. -
2 2 ..
1 ..
n 1 ..
r. ,I1 ! : : A. n
Fig. 8.7 Histograms of Cibicidoides alleni and Subbotina patagonica before (A) and after (B) iteration.
After iteration, the bimodal histogram of C. alleni has remained approximately the same, whereas the
histogram of S . patagonica has become very narrow.
290

The tail of the frequency distribution then extends in the stratigraphically

upward direction with predicted positive skewness of the frequency
distribution (cf. Fig. 2.10D).
The skewness of the histograms for 44 Cenozoic foraminifers along
the northwestern Atlantic Margin has been determined by computing
their (unbiased) sample skewness statistics (see Table 8.16). (The
“unbiased” skewness was obtained by multiplying the sum of cubes of
standardized deviations from the mean by nln-l)(n-2)). In column3 of
Table 8.16 the skewness was estimated for deviations from the best-fitting
spline-curves. Although individual estimates of skewness are not
significantly different from zero ( = symmetry), because sample sizes are
small (from 7 to 22 only), column 3 shows a pattern in that the events in
the upper half of the table display almost exclusively negative values for
skewness, whereas those in the lower half are almost all positive. This
pattern partly can be explained by the fact that RASC distances near the
tops of the sections were underestimated whereas those near the bottoms
were overestimated (cf. Fig. 8.6). Bias introduced by use of estimated
means which are too low or too high can be eliminated by substituting the
mean deviations plotted in Figure8.6B for the sample mean in the
equation used for estimating skewness. The resulting revised estimates
are shown in column 4 of Table 8.16. Clearly skewness was increased near
the top of this table and decreased near its bottom. However, the pattern
remains that in the upper half of the table, most skewnesses are negative,
whereas those in the lower half are mostly positive. It is noted, that 6 of 8
species a t the bottom of the table have negative skewness in column 4 of
Table 8.16.
Comparison of the RASC distance scale to the geological time scale
shows that the positive skewness values are largely restricted t o the
Eocene which extends approximately from event 56 t o event 259 (cf.
Gradstein et al., p. 339) corresponding to a time interval of about 2 1 Ma
(from 58 t o 37 Ma). The total range of RASC distances in Tables 8.15 and

TABLE 8.16

Selected statistics for the 44 species after convergence. Degrees of freedom f,= ni-1 where ni represents
sample size for event i. Skewness 1 and 2 are sample statistics per species using zero mean and sample
mean for deviations from spline-curves, respectively. The pooled variance s2 is equal to 0.351.
Variance ratio s,2/s2 has asterisk if its value is below 0.005 fractile or above 0.995 fractile of
corresponding x 2 / f distribution. Last column shows individual terms added to give Bartlett’s
9 2 = 180.734 (see text). Constant C= 1.034 was computed by formula in Hald (1975, p. 291).
Event h Skewness 1 Skewness2 sz,/sz f , * h ( S ~ ~ I ISC~ )
10 9 -1.367 -0.059 3 900' -9.589
17 11 -1.678 -1.276 1 999 -7.367
16 21 -1.392 0.205 0 745 5.983
67 7 -2.375 -1.297 1 492 -2.710
18 21 -1.140 -0.451 0 264 27.034
21 9 -1.074 -0.507 0 025; 32.066
20 19 -1.542 -1.108 0 198' 29.681
71 12 -1.040 -0.617 1061 -0.683
26 12 -0.016 0.368 1172 -1.838
70 6 -0.479 -0.965 0 384 5.556
15 21 -1.548 -1.284 1 I92 -3.570
24 16 -0.792 -0.469 0 512 10.370
27 12 -1.313 - 1 045 2 094 -8.575
69 10 -1.139 -0.253 1 799 -5.680
25 18 -0.586 0.233 0 677 6.778
81 11 -1.652 -0.563 1 776 -6.109
202 6 -1.499 -1.153 0 266 7.689
259 13 -0.357 0.495 0 263' 16.782
147 6 -0.812 0.601 0 472 4.359
34 14 -0.727 0.103 1578 -6.172
33 6 -0.404 0.148 3 251* -6.841
260 14 1.681 1.442 0 431 11.399
261 14 1.920 0.809 0 199' 21.836
263 12 0.791 0.425 0 998 0.038
29 18 -0.034 -0.027 0 385 16.633
32 17 -0.481 .0.836 0 627 7.672
40 9 1.207 0 651 1 232 -1.816
42 12 1.356 0.859 2 399. -10.I57
264 6 2.403 1 808 1023 -0.131
41 11 0.358 0.429 1 029 -0.307
30 11 0.600 0 229 1185 - 1 816

90 6 1 084 1.894 1175 -0.936

3fi 10 0511 0 424 1072 -0.676
8fi 6 0.890 0 271 0 093' 13.789

57 18 0 469 0.150 1 586 -8.030

45 9 1511 0.185 2 634' -8.429
50 8 0 118 .1.394 0 006; 39.602
46 13 1.361 0.038 1119 -1.414
230 6 1.466 -0.675 1124 -0.677
54 12 1.659 0.573 0 632 5.334
52 6 -0 333 - 1.424 0 478 4.285
56 13 1.486 -0.046 0 539 7.764
55 8 1.388 -1.278 0 790 1.821
59 7 1321 .1.597 1465 -2 587
292

8.16 corresponds to about 63Ma. The species with positive skewness,

therefore, tend to occur during the epoch (Eocene) that is represented by
relatively many species in our application. It seems t h a t M o d e l A
predominated during this time interval, whereas Model B predominated
after and possibly before the Eocene. This result is corroborated by the
observation that tests usually are reworked in the younger Neogene
section of the Labrador Shelf (cf. Section 4.7).
It was assumed in the previous section that variances si2 obtained for
the species are significantly different from one another. This assumption
has been tested statistically with the results shown in the last two columns
of Table 8.16. Column 5 shows species variances si2 divided by s2 = 0.351
representing the pooled variance for all 44 species (see before). If the
variances are equal, this ratio is approximately distributed as x2/f= .s2/a2
where the chi-squared (x2) has fdegrees of freedom. The fractiles of this
distribution have been tabulated for different values of f by Hald (1960,
p. 44). In Table 8.16, an asterisk was given t o values below the 0.005 or
above the 0.995 fractile. Such values would occur with probability
a = 0.01. This test indicates that six variances are probably too small and
four are too large in Table 8.16. Bartlett’s x2-test for equality of variances
(see e.g. Hald, 1957, p. 291) has also been applied. According t o this test,
the quantities in the last column of Table 8.16 would add up to x2 with (k-
1) = 43 degrees of freedom. The total chi-squared value is equal to
180.734 which far exceeds the corresponding 99% confidence limit
(= 67.5). Bartlett’s chi-squared test, therefore, also indicates that the
variances si 2 are not equal t o one another.
Another statistical experiment conducted for this example is as
follows. From the preceding results, it may be concluded t h a t the
variances of the 44 species are not equal to one another. For this reason,
the values used for the histograms of individual species were standardized
by dividing them by S i . Consequently, 44 sets of values were obtained with
means equal to zero and standard deviations equal to one. These 44 sets of
values were combined with one another t o give a single new set of 550
standardized values of which the histogram is shown in Figure 8.8. This
composite frequency distribution would be positively or negatively skew if
the frequency distributions for individual species would all tend to be
asymmetric, e.g. according to Model A or B (see before). Instead of this, the
composite distribution (Fig. 8.8) seems to be approximately symmetric.
When the last two classes in upper and lower tail are combined with each
other, 13 observed frequencies are retained for the histogram of Figure 8.8
 293

-2 6 -1 8 -1

Standardized deviations

Fig. 8.8 Histogram of 550 standardized differences from all spline-curves for all species after
convergence. Standardization was achieved by dividing each difference by the standard deviation sL for
its species.

which can be compared to 13 theoretical frequencies obtained from the

normal distribution in standard form. Application of the chi-squared test
for goodness of fit gave ?2(10) = 12.03 for the difference between observed
and theoretical normal distribution. For 10 degrees of freedom, the
corresponding 95% and 99% fractiles of the x2-distribution are 18.3 and
23.2, respectively. Because the jj2-value estimated for Figure 8.8 is less
than these values, it may be concluded that the composite distribution of
Figure 8.8 is approximately normal (Gaussian).
Earlier in this section, positive and negative skewness of individual
frequency distributions was discussed. Although sample sizes are too
small t o establish that the individual skewness values of Table 8.16 are
significantly different from zero, the sign of skewness changed through
time according to a regular (nonrandom) pattern. Obviously, this pattern
is too weak to show up as a systematic departure from normality in the
composite frequency distribution of Figure 8.8.
294

The modified RASC method consists of alternately obtaining two

different estimates ( x i and 32,) of the mean position EX, of each event i
along the relative time interval scale. This iterative process converges t o a
final solution which does not differ greatly from the ordinary RASC scaled
optimum sequence. The differences (32,-3,) provide a n estimate of the
frequency distribution for event i. It has been demonstrated that the
highest occurrences of Cenozoic Foraminifera along the northwestern
Atlantic margin have different variances. The histogram of standardized
values for all species was shown t o be approximately normal. The
possibility t o identify good markers with small variance (e.g. Subbotina
putugonica) is a new feature of modified RASC not previously provided by
ordinary RASC. Likewise, it has become possible to identify relatively
poor markers with relatively large variance and perhaps bimodal
distribution (e.g. Cibicidoides alleni). Although xi and fi both provide good
approximations of EXi, some bias was introduced during the iterative
process consisting of reduction of average variance as well as non-zero
mean values of (32i-xi) for events near top and bottom of the stratigraphic
sequence. The method also provides a way t o construct conservative range
charts in which the ranges of the fossils are extended to the highest
occurrences in individual sections.
For example, in Figure 8.7B, the largest (positive) deviations on the
right side of the frequency curves are plotted at 0.1 and 1.7, respectively.
These values can be added to the RASC distances (sixth run, Table 8.14) in
order t o obtain conservative ranges. (The maximum positive deviation
exceeded 1.5 for only two of the 550 values used in the histograms for
separate events. In these two situations, the range extension was set
equal t o 1.7). Figure 8.7 shows highest occurrences based on cumulative
(modified) RASC distances (A) a s well as highest occurrences for
individual sections (C) obtained by subtracting the largest positive
deviations. For comparison, the mean deviations (B) of Figure 8.6B also
are shown in Figure 8.9 in the form of positive or negative deviations from
the RASC distance (A).
If all variances were equal to 0.5,95percent of the positive deviations
would be less than 1.163. This was the value previously used for the range
extensions in the Drobne example of Figure 8.4. It was shown by analysis
of variance that the variances of the taxa in the Gradstein-Thomas
database are not equal t o one another. Thus the shorter range extensions
in Figure 8.9 are for taxa with variances which are significantly less than
the average variance. On the other hand, it should be kept in mind that
 295

1 .o

2.0

:3.0

I
.-
m
1

U m
h

u
vI 4.0
2 I?
5.0

6.0

7.0

Highest occurrences in order of estimated RASC distance(A)

Fig. 8.9 Extended RASC ranges for Cenozoic Foraminifera in Gradstein-Thomas database. Letters for
taxon 59 on the right represent (A) estimated RASC distance, (B) mean deviation from spline-curve, and
(C) highest occurrence of species (i.e. maximum deviation from spline-curve). B is shown only if it differs
from A. Good markers such a s taxon 50 (Subbotinaputugonica)have approximately coinciding positions
for A, B and C. Note that a s a first approximation it could be assumed that the highest occurrences (C)
have RASC distances which are about 1.16 units less than the average position (cf. Section 8.3). This
systematic difference in distance is equivalent to approximately 10 m.y. (cf. Fig. 9.2, see later).

the range extensions have their own variances and are subject to more
uncertainty t h a n t h e RASC distances themselves. The subject of
conservative range charts also will be discussed in the next two sections
with applications to smaller datasets.

8.8 Application of modified RASC to Drobne’s alveolinids

The Drobne example (cf. Section 8.3) was subjected to modified RASC
instead of RASC with results shown in Tables 8.17 and 8.18. Sections V,
IX and XI have only one or two event levels (see Fig. 8.1) and could not be
used in modified RASC because at least 3 event levels are needed for
curve-fitting. The scaled optimum sequence previously obtained by RASC
TABLE 8.17

Modified RASC method applied to original Drobne example of Section 8.3. After 4 iterations, the RASC
distances ($4) are close to the original RASC distances ($1). The event variances ( 9 4 ) are for zero mean
deviations and differ from one another. Degrees of freedom (d.f.) in last column are equal to 3 or 4 for
nearly all events. For 3 degrees of freedom the 95% confidence interval of the sample variance ranges
from 0.3202 to 3.1202. H e r e 4 is the expected value of the variance which is approximately equal to 0.5 in
this application. According to this single variance test, the variance of event 15 would be too large and
those of events 20,27,22,2,23, 1 and 3 would be smaller than average. However, modified RASC gives
results that are approximate if samples sizes are very small. It will be seen later (see Table 8.21) that
only the variances of events 27,2 and 1 are again much smaller than average after enlarging the dataset
and re-running modified RASC.

Event X1 r4 SP4 d.f.

28 0.00 0.00 0.31 4
20 0.02 0.11 0.05 4
19 0.30 0.32 0.14 4
18 0.45 0.45 0.45 3
27 0.88 0.76 0.06 4
15 1.16 1.16 3.04 3
17 2.00 2.02 0.76 3
22 2.02 2.07 0.07 3
2 2.16 2.20 0.03 3
23 2.16 2.18 0.05 4
21 2.32 2.33 0.26 3
1 2.47 2.45 0.13 3
14 2.69 2.69 0.30 6
12 2.70 2.70 0.26 4
25 2.89 2.89 0.33 4
11 3.33 3.33 0.44 4
5 3.33 3.32 0.96 3
13 3.52 3.53 0.43 6
3 4.60 4.60 0.00 3

is shown as 51 in Table 8.17. It was the starting point for modified RASC
which, after four iterations, produced nearly the same scaled optimum
sequence ( f 4 in Table 8.17).
It is noted that on the basis of the results by modified RASC described
in the previous section (also see D’Iorio, 1988) indicating that the order of
events does not change significantly when this method is applied, it was
 297
TABLE 8.18

Deviations of observed relative positions of events from spline-curves after 4 iterations. Numbers along
top indicate the eight sections used. Event numbers are given in first column. Events 15,23,25,5 and 3
have asterisk for coinciding highest and lowest occurrences in all sections. The variances of Table 8.17
were based on these numbers. Largest deviations for even code numbers (=highest occurrences) and
lowest deviations for odd code numbers (=lowest occurrences) were used for range chart of Fig. 8.10.
These numbers are shown in bold print. Rows with asterisks have two bold numbers.

1 2 3 4 6 7 8 10
28 X -0.97 -0.23 -0.04 -0.47 -0.07 X

20 X X -0.12 0.07 -0.37 0.04 -0.22

19 X X 0.08 -0.68 -0.16 X X

18 X -0.52 0.21 0.40 -0.93 -0.21 X

27 X -0.20 -0.19 -0.23 0.29 -0.21 X

15* -0.98 X X -0.78 X 0.18 2.74

17 X -0.09 1.06 -0.86 0.64 X

22 X -0.03 0.39 0.13 X -0.17

2 X 0.10 X 0.26 -0.07 -0.05
23* -0.27 0.08 -0.23 0.24 X -0.06
21 X 0.23 0.64 -0.55 X 0.09
1 X 0.34 X -0.44 0.17 X X 0.20
14 0.24 0.59 -0.45 -0.19 0.42 -0.04 -1.00 X

12 -0.54 0.60 -0.44 X X X -0.08 0.46

25* X -0.34 -0.25 X -0.28 0.16 1.01 X

11 0.08 0.09 -0.54 X X X 0.54 1.08

5* 1.19 -1.04 -0.54 X X -0.34 X X

13 1.08 -0.83 -0.34 0.65 0.36 -0.13 -0.16 X

3* 0.00 X 0.01 X X 0.00 0.00 X

decided to change the procedure slightly as follows. Instead of taking the

scaled optimum sequence without final reordering as the starting point, it
is now possible to take the scaled optimum sequence after final reordering
as the starting point. On the other hand, the order of events is not allowed
to change during successive iterations in modified RASC. The order of
events in 34 in Table 8.17 is identical to that in f l except for events 11 and
5 which are nearly coeval on the average.
298

The variances of the events (s24) had not completely converged after 4
iterations. Because the number of degrees of freedom for s24 is small for
all events ranging from 3 to 6, these results are subject to considerable
uncertainty. According to Table 8.17, events 2 and 3, corresponding to the
highest occurrence of species 1 (A. moussoulensis) and the lowest
occurrence of species 2 (A. aramaea) have variances closest t o zero and
could be good marker horizons. However, these two events each occur in 4
sections only. The fact that their positions are on the fitted spline-curves
may not be significant because there are so few data. It should be kept in
mind that small variance events receive relatively more weight than other
events in spline-curve fitting. In fact, zero-variance events have the
property (cf. Section 3.11) that the best-fitting spline-curve is forced to
pass exactly through their points on the scattergram. The possibility,
therefore, exists that an event which happens t o have a small variance
because it occurs in so few sections, obtains zero-variance during the
convergence process which involves repeated spline-curve fitting for all
sections.
The final deviations of the 19 events from the 8 fitted spline-curves
are shown in Table 8.18. If all variances are assumed to be equal, numbers
with absolute value greater than 1.16 denote events out of position with
probability greater than 95%. The two events with this property are event
15 (species 8) and event 5 (species 3). The latter event occurs in a reworked
bed as discussed in Section 8.3. According to the preceding equal variance
test applied to Table 8.18, species 8 would occur too high in Section X.
However, this result would need confirmation by additional evidence or
other experiments because there are too few event levels per section in this
dataset for a fully convincing application of modified RASC.

Brower (1990) has carried out a method comparison study on the

Drobne dataset. Figure 8.10 shows ranges for 12 species obtained by 5
methods. The ranges resulting from the Unitary Associations (U.A.)
method, seriation (SER) and RASC were calculated by Brower and plotted
along a relative time-scale with 10 units. The RASC distances 4 of Table
8.17 were enlarged by the factor (10/4.16=) 2.40 so that their largest value
(for lowest occurrence of event 2) became 10 instead of 4.16 in Table 8.17.
These RASC distances are shown as tick marks on the left of the ranges for
each species in Figure 8.10. Species with coinciding highest and lowest
occurrence in all sections have a single tick mark only.
 299

Fig. 8.10 Comparison of five types of ranges for Drobne’s alveonilids along relative time scale of Brower
(1990) who pointed out that RASC ranges are significantly shorter than Unitary Associations (U.A.) and
Seriation (SER) ranges. These results are compared to the modified RASC (MR) ranges and the average
highest occurrences (ave HO) and average lowest occurrences (ave LO) on which these MR ranges are
based. The relative time scales used for U.A., SER, RASC and MR, respectively, have different units and
are not completely comparable (cf. Brower, 1990). However, on the whole, the MR ranges are about as
wide as the U.A. and SER ranges.

The ranges between tick marks were extended by adding deviations

from Table 8.18 as follows. For highest occurrences (even numbers in
Table 8.18), the largest deviation was subtracted from the RASC distance;
for lowest occurrences, the absolute value of the smallest deviation was
added to the RASC distance; and for species with coinciding highest and
lowest occurrence, both the largest and the smallest deviations were used.
The resulting extended ranges are shown in Figure 8.10.

Brower (1990) used his own computer algorithms for U.A. and RASC
which differ somewhat from those used by Davaud and Guex (1984) and in
Gradstein et al. (1985). Also, because different methods have different
time-scales, plotting all ranges along a single time-scale may distort some
300

results. However, Brower (1990) correctly concluded that the average

ranges obtained by RASC were significantly shorter than the ranges
obtained by U.A. and seriation. The distances between ave HO and ave
LO are very close t o the Brower’s RASC ranges, and the extended modified
RASC (MR) ranges are approximately as wide as the U.A. and SER
ranges. For species 8, 9 and 3, the MR ranges are wider than the other
ranges. These wider extensions are in part due t o the “anomalous” values
(greater than 1.16)for species 8 and 3.
The number of event levels per section can be enlarged by not using
the maximal horizons method for data reduction. Table 8.19 is based on
use of all stratigraphic information on relative positions of highest and
lowest occurrences. For example, Section I1 (2) for Figure 8.2 has 9 event
levels in Table 8.19 versus 4 maximal horizons in Figure 8.1. The
reworked bed (level 4 in Section I of Fig. 8.1) was not included in the SEQ
file of Table 8.19. The new scaled optimum sequence obtained after final
reordering is shown as 31 in Table 8.21.

Table 8.20 shows normality test results for the 3 sections with events
that are anomalous with a probability of 99%(2 asterisks for second-order

TABLE 8.19
SEQ tile for recoded Drobne dataset. Most sections have more event levels than in Fig. 8.1. Section 2
(Dane near Divafa, see Fig. 8.2) has 9 event levels which were reduced to 4 maximal horizons in Fig. 8.1.
The number - 999 denotes end of section in SEQ file.

SECTION 1
15 -16 7 -8 -13 -14 -23 -24 11 -12 3 -4-999 0 0 0 0 0 0 0
SECTION 2
28 18 -21 2 -14 -24 1 - 1 2 -17 -21 -22 23 11 -25 -26 4 -6 15 -16 3
-5 -9 -10 -13-999 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
SECTION 3
18 -20 28 19 30 27 17 21 -22 23 -24 -29 14 -26 12 -25 6 -11 5 -13
3-4-999 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
SECTION 4
20 -28 18 29 -30 7 -8 -19 -27 2 -15 -16 -22 -23 -24 1 - 1 3 -14 -17 -21
-999 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
SECTION 5
7 -8 9 -10-999
SECTION 6
19 -20 -27 -28 7 -8 -15 -16 1 -2 -14 13 -25 -26-999 0 0 0 0 0
SECTION 7
14 -25 -26 -29 -30 5 -6 -13 9 -10 3 -4-999 0 0 0 0 0 0 0
SECTION 8
20 -28 19 15 -16 -27 11 -12 -25 -26 13 -14 4 -10 3 9-999 0 0 0
SECTION 10
23 15 -16 19 -20 24 1 -2 -11 -12 -21 -22-999 0 0 0 0 0 0 0
SECTION 11
19 -20 -27 -28 7 -8 -17 -18 1 -2 -14 13 -25 -26-999 0 0 0 0 0
 301
TABLE 8.20

RASC normality test output for the 3 sections in the recoded Drobne dataset with one or more events
with double asterisks.

SECTlON 1 CUM. D l S r . 7ND ORDER DIFF.

LO A . LEUPOLDI 1) 1.9144
HI A . I.EUI'0LUl -I6 1.914' - 1.1814
LO A . GLOBOSA I I . 3920 1. '3814
HI A . GLOHOSA -8 I . 397.0 %.3005 9:
I,0 A . PASTICII.IATA - 11 3.6925 - 1 ,1,991, ?:9:
111 A . PASTICILLATA - I4 2.11935 0.61432
LO A . SUBPYRENEICA -21 1.9371 0.4907,
HI A . SUBPYRENEICA -24 1.8122 0.5950
LO A. PlSlFORMlS 11 '3 .260', -0.9526
HI A . PISIFOKMIS -12 2 ,533 7 4 I. 7861,
10
. A . MUMAFA 1 5.0S96 -I.9?l7
HI A . MUMAFA -4 4.491%

SECTION 2 CUM. IIIST. 2ND OMIER DIFF.

HI A . GUIDONIS 28 0.0000
HI A . WNTANAKII 18 0.5241 0. 10'32
I,0 A. G U I W N I S -17 0.6910 0.6962
HI A . MOlISSOUI,F.NSIS 7 7.0151 -0.3842
HI A . PASTICILLATA -14 2.4935 -1.0991
I11 A . SUBPYRENEICA -7.4 1.8722 O.H391
10
. A . MOUSSOIJLENSIS I 7.5'117 0.0678
HI A. PISIFORMIS - 12 7. .83 11, -1.1310
LO A . MOEPTANARII -17 1.9921 1.2539
LO A . UELW)LIA - 7.1 7.. 4006 -0.1461
HI A. IIEDULlA -22 2.0631 -0.1494
L,O A . S W P Y R E N E I C A 23 1.9377 1.4482
1.0 A . PISIFOKMIS 11 3.2605 -n.9277
LO A . wu(A -25 '3.1941 -0.1121
lil A . lA4.4 -26 1.01,,6 I . 1926
HI A . ARAMAFA l4 .
4 1,') 17 -4.0524 ??::

1.0 A . LEUPOLL11 15 1.9144 '\,"383 ;R

HI A. 1.EIJPOLUI - I6 1.9141~ 2.6836 ?:

LO A . A W A 3 5.05'>6 -2.8790 f

LO A . AVEI.IANA -9 1, ,8642 -0.2161

ti1 A . AVELLANA - 10 11.4521 -0.3481
1.0 A . PASTICILLATA -11 3.6925

SECTION n WM. IIIST. 2ND OKUER DIFF.

HI A. AKACONENSIS 20 O.l?61
HI A. GUIWNIS -78 o.ooon 0.1094
1.0 A. AKACONENSIS 19 0.6595 (1.5955
1.0 A . I.I'uP0LDI I5 1.9144 -0.5782
HI A. LKUP0I.DI -16 1.9144 -1.2234
1.0 A . GUlUONlS -? 1 11.6910 3.1161 ;S':
1.0 A. PISIFOKMIS 11 3.2605 -2.3158 f
HI A. PlSIFflRMIS - 12 2.8374 0.1798
1.0 A . IAXA -7 5 1.1941 -0.5352
HI A. w(A -26 3.0156 n. i i n 6
LO A. PASTICIISATA I3 '3 .6925 -1.1991
HI A. PASTICILIATA -14 2.4931 7 . 5199 <:
HI A. AKAMAEA 4 4 ,4912 -1.3594
HI A. AVELSANA - 10 4 .4 5 2 1 -0.031~
1.0 A . ARAMAFA 3 5 . 0 5 ~ -0.8023
1
.0 A . AVELLANA > 4.864?
302

TABLE 8.21

Modified RASC method applied to recoded Drobne dataset. n is number of sections in which event was
observed. f 1 , i 3 and f 4 are RASC distances at beginning and after 3 and 4 iterations, respectively.
Variances after 3 and 4 iterations are for zero mean deviation and are only approximately equal to one
another. SK1and SK2 are skewness statistics with and without zero means, respectively.

Event n fl f3 n4 s23 s24 SK 1 5k2

28 6 0.00 0.00 0.00 0.09 0.12 -1.61 0.30
20 6 0.13 0.12 0.13 0.41 0.42 -2.70 -1.55
18 4 0.52 0.51 0.52 0.43 0.45 -0.86 1.18
19 6 0.66 0.64 0.64 0.29 0.29 -1.83 -0.04
27 6 0.69 0.67 0.68 0.46 0.49 -2.13 -0.52
8 5 1.39 1.37 1.38 0.03 0.01 -1.57 -1 84

7 5 1.39 1.37 1.38 0.03 0.01 -1.53 -1.85

24 5 1.87 1.84 183 0.05 0.07 -1.77 -1.12
16 6 1.91 1.86 1.86 0.99 0.92 -2.28 -1.97
15 6 1.91 1.86 1.86 0.99 0.92 -2.28 -1.97
23 5 1.94 1.88 1.88 0.25 0.27 -0.91 -0.30
17 4 1.99 1.94 1.99 0.35 0.34 0.56 0.01
2 5 2.02 2.02 2.01 0.04 0.04 -0.12 0.60
22 4 2 06 2.00 2.00 0.06 0.05 -1.72 -0.08
21 4 2.40 2.27 2.27 0.07 0.06 -2.08 0.92
14 8 2.49 2.41 2.41 0.37 0.40 -0.53 -0.51
1 5 2.55 2.42 2.42 0.05 0.05 1.76 -1.39
12 5 2.84 2.72 2.72 0.15 0.14 1.15 0.07
26 6 3.01 2.91 2.91 0.03 0.02 1.15 0.94
25 6 3 20 3.02 3.02 0.02 0.02 2.84 1.76
11 5 3.26 3.09 3.08 0.28 0.28 1.84 -0.03
13 8 3.69 3.54 3.54 0.83 0.89 1.95 0.77
10 4 4.45 4.30 4.29 0.00 0.00 -2.17 -1.13
4 5 4.49 4.35 4 35 0.34 0.35 1.37 1.02
9 4 4.87 4.71 4.71 0.64 0.59 1.66 -2.00

3 5 5.06 4.92 4.92 0.24 0.23 2.05 -0.11

differences). The highest and lowest occurrence of species 7 (A.

pasticillata) coincide in Section I(1). The lowest occurrence occurs too high
in this section in comparison with its neighbors. On the contrary, species 8
 303

( A . leupoldi), of which the highest and lowest occurrence coincide in all

sections containing it, occurs too low in Section I1 (2). This is not
immediately obvious from the pattern of asterisks for this section but
follows when it is considered that its highest and lowest occurrence have
the same cumulative RASC distance. Finally, the lowest occurrence of A .
guidonis may be situated too low in Section VIII (8). The suggestion on the
basis of Table 8.18 that A. leupoldi occurred too high in Section X (10)is
not confirmed by the new normality test results.
Table 8.21 shows modified RASC results. The RASC distances (54)
after four iterations are nearly equal to those (23) after three iterations.
Comparison to the original scaled optimum sequence (fl) for t h i s
experiment shows that the modified RASC method left the scaled optimum
sequence nearly unchanged. The variances (unbiased estimates of
deviations from the spline-curves) have not yet fully stabilized after 4
iterations because the values of s24 differ from those of s23 in Table 8.21.
Slightly more events have negative skewness after 4 iterations but there is
no systematic pattern of change in the skewness. There is no evidence
supporting prevalence of either Model A or Model B during a particular
time interval (cf. Section 8.7).
Extended modified RASC ranges for the newly coded Drobne dataset
are shown in Figure 8.11. The three events identified a s possibly
anomalous by the normality test (Table 8.20) had deviations from the
spline-curves exceeding 1.16 representing the 95% confidence limit if all
event variances are equal. This indicates good agreement between
normality test and modified RASC results.
Elimination of anomalous events would shorten the extended ranges
for species 14 and 8 by the amounts shown in Figure 8.11. The assumption
that the lowest occurrence of species 7 occurs too high in Section I (cf.
Table 8.20) does not change the length of the extended range for this
species. Because the modified RASC range chart of Figure 8.11 is based on
more information than the corresponding range chart of Figure 8.10 it is
probably better. It is not possible to determine by how much the procedure
of recoding the Drobne data followed by modified RASC has improved
upon the original RASC extended range chart shown in Figure 8.3. The
new result is closer t o Drobne’s subjective zonation. It also is interesting to
recall Guex’s remark (see Section 8.3) that species 5 probably never
304

-1 .o

1 4 10 9
0.0
T
1 .o
1
1 4 a
W
7
V
C
(0 2.0 6
e
.-ul
'0 T 13
$ 3.0
d
s

4.0

t ave HO

5.0
-
.-ave LO
LO

6.0
Species with code numbers

Fig. 8.11 Extended modified RASC ranges for Drobne's dataset. As in Fig. 8.3, the species were ordered
on the basis of the RASC distances of their average highest occurrences. The sample sizes were small
and this is the main reason for the random fluctuations in the positions of the highest (HO) and lowest
(LO) occurrences. Deletion of events with double asterisks in normality test (see Table 8.20) would result
in shorter ranges for species 8 and 14 as shown by arrows in Fig. 8.11.

coexisted with 11 although their ranges overlapped in Figure 8.3. In

Figure 8.11, the extended ranges for these two sepcies are clearly separate.
It is good t o keep in mind that the highest and lowest positions of
fossils in individual sections are subject to more uncertainty than the
positions of the average highest and lowest occurrences (cf. Chapter 2). In
general, it is better to base the construction of isochrons in automated
stratigraphic correlation (see Chapter 9) on average highest and lowest
occurrences because these can be known with more precision than the
(conservative) truly highest and lowest occurrences. On the other hand, if
assemblages of fossils are used for subjective correlation, the extended
range chart may provide a better tool than a range chart which is based on
average stratigraphic events.
 305

8.9 Comparison of range charts for Palmer’s database

Application of the normality test t o Palmer’s (1954) database for the
Riley Formation in central Texas was discussed in Section 8.4. This
section contains results of the modified RASC method for this example.
The scaled optimum sequence shown in the third column of Table 8.22 was
taken as the starting point for a run with seven iterations in total.
Approximate convergence was obtained as shown in Table 8.22. Figure
8.12 shows the final spline-curve in comparison with observations for the
Morgan Creek section. This is output from the modified RASC module of
micro-RASC (cf. Chapter 10)
Table 8.23 contains deviations used for the extended range chart with
lowest and highest occurrences for each taxon. The deviations between
points and curve graphically shown in Figure 8.12 for the Morgan Creek
section correspond t o the deviations listed in the column numbered 1 in
Table 8.23. The mean deviation per event (in Table 8.22) was subtracted
from the original deviations between points and curves before entering
them into Table 8.23.
The extended ranges are shown in Figure 8.13 for comparison with 4
other range charts. The central 3 sets of ranges were taken from Edwards
(1982) who used the same abbreviated dataset of 16 taxa representing
those taxa occurring in 5 , 6 or 7 sections. Edwards’ comparison was for (1)
Shaw’s (1964) final composite standard, (2) Edwards’ (1978) conservative
method, and (3) Hay’s (1972) method as applied by Edwards. The 32
lowest and highest occurrences were ranked from 1 to 32 for each method.
The ranges obtained by the first two methods were considerably wider
than those obtained by the Hay method which is comparable to RASC.

The RASC ranges plotted in Figure 8.13 are the final modified RASC
distances ( i 7 ) of Table 8.22. Deviations for highest and lowest occurrences
were taken from Table 8.23. For most taxa in Figure 8.13, the three
ranges on the left side (modified RASC, Shaw and Edwards ranges) are
approximately equally wide. The same holds true for the two ranges on
the right (Hay and RASC ranges) which are considerably shorter than the
three ranges on the left. On the whole, modified RASC has the widest
ranges, partly because its ranges are clearly wider than the Shaw ranges
for taxa 4, and 36 (with highest occurrences 8 and 70, respectively). The
deviations in Table 8.22 corresponding to these two taxa are 1.27 (for event
8) in the Morgan Creek section and -1.07 (for event 69) in the Pontotoc
306

TABLE 8.22

Modified RASC method applied to Palmer’s database. Approximate convergence was reached after 7
iterations. See Table 8.21 for explanations of column headings. The average deviation (Ave) is
significantly less than zero for the first 6 events listed (see text for further discussion).

Event n XI ffi P7 sls dfi 61, Ave SKI SK2

100 I 0 00 0 00 0 00 1 73 1 76 174 112 I53 0 66

82 I 0 31 0 23 0 23 1 43 145 I 43 0 97 161 0 53

108 5 0 46 0 42 0 42 I 39 142 138 0 99 168 0 77

107 5 132 1 24 1 24 0 66 0 67 0 65 0 59 1 94 0 44

99 7 146 138 1 38 0 64 0 64 0 64 0 49 -2 05 0 54
90 5 2 02 I 95 1 95 0 55 0 55 0 55 0 54 2 13 0 59
92 6 2 17 2 40 2 40 0 02 0 02 0 02 0 08 2 06 0 32
89 5 2 82 2 83 2 82 0 07 0 07 0 07 0 06 0 44 0 80

91 6 2 85 2 89 2 88 0 06 0 07 0 07 0 04 0 90 0 28
81 7 3 08 3 11 3 10 0 07 0 07 0 07 0 08 1 28 0 16

68 I 3 86 3 63 3 62 0 08 0 08 0 08 0 05 0 72 0 02
70 6 3 99 3 80 3 80 0 14 0 15 0 15 0 13 0 11 1 84

42 7 4 02 3 81 3 81 0 18 0 18 0 19 0 10 0 36 0 60
69 6 4 28 4 14 4 14 0 57 0 51 0 59 0 12 0 19 0 88
24 5 4 35 4 21 4 21 0 30 0 30 0 30 0 28 I 17 0 43
61 7 4 51 4 40 4 42 0 23 0 19 0 20 0 21 129 0 52
40 I 4 69 4 55 4 55 0 93 0 91 0 91 0 19 0 64 I55
56 5 4 71 4 64 4 64 0 66 0 63 0 63 0 10 1 16 0 53
54 5 5 05 4 91 4 91 0 20 0 19 0 19 0 II 0 96 0 20

48 5 5 21 5 07 5 07 1 44 1 40 1 39 0 23 127 0 36
47 5 5 21 5 07 5 07 I 44 140 I39 0 23 1 27 0 36
55 5 5 46 5 37 5 36 0 12 0 12 0 12 0 05 0 66 I 39
53 5 5 47 5 33 5 33 0 20 0 20 0 20 0 02 0 25 0 03
41 7 5 62 5 45 5 44 1 27 I24 1 23 0 13 I 48 1 02

34 I 5 64 5 49 5 46 0 08 0 07 0 07 0 10 1 I4 0 33
22 6 5 75 5 61 5 61 0 17 0 17 0 17 0 02 0 53 0 26
8 I 5 82 5 63 5 63 0 98 0 97 0 96 n 06 0 70 0 14
23 5 6 10 5 91 5 90 0 25 0 25 0 25 0 10 I 04 0 04
39 7 6 13 5 91 5 90 0 16 0 15 0 14 0 07 2 36 I70

33 7 6 26 6 04 6 03 0 17 0 I6 0 I6 0 13 2 35 I 33
21 6 6 I9 6 53 6 52 0 II 0 11 0 II 0 18 2 28 0 80
7 I 6 85 6 68 6 69 0 41 0 46 0 47 0 21 I 96 0 94

section, respectively. In absolute value, these two numbers are close t o

1.16, representing the 95% confidence limit for anomalous values if all
event variances are equal. Shaw (1964) did not use these events for
 307
MORGAN CREEK

R
A
S
C
D
I
S
T
A
N
C
E

Fig. 8.12 Comparison of observed highest and lowest occurrences (shown as x-es) with best-fitting spline-
curve (=straight line) after iteration. The line shows a relatively poor fit at the first two event levels.
The RASC distances plotted in the vertical direction are close to those of the scaled optimum sequence
used in Fig. 9.23(see later). The spline-curve in Fig. 9.23was obtained by cross-validation and provides
a better fit than the straight-line fit of Fig. 8.12.

constructing his composite range chart. However, neither event was

flagged as possibly anomalous in the normality test (Table 8.13).
The largest deviation ( = 1.60) in Table 8.23 is for event 41 in the
Pontotoc section. The corresponding second-order difference has two
asterisks in Table 8.13 suggesting a possible anomaly. However, because
TABLE 8.23

Deviations of observed relative positions of events after 7 iterations. Values were corrected for average
deviation from spline-curve (Ave in Table 8.22). Numbers 1 to 7 for columns correspond to the 7 sections.
Largest deviations in bold print were used to construct modified RASC range chart of Fig. 8.13.

Event n 1 2 3 4 5 6 7
~

100 7 -0.41 0.34 -0.05 -0.49 0.26 0.89 -0.53

82 7 -0.66 -0.38 -0.57 0.04 0.34 0.97 0.26

108 5 -0.12 0.23 -0.36 X 0.55 -0.31 X

107 5 -0.35 0.66 -0.52 X 0.09 0.12 X

99 7 -0.32 -0.11 -1.07 0.26 0.13 0.89 0.22

90 5 0.31 -0.29 0.13 0.43 X X -0.58

92 6 -0.03 0.10 X -0.03 -0.15 -0.00 0.12

89 5 -0.27 0.11 -0.04 0.38 X X -0.18

91 6 -0.31 0.06 X -0.11 0 22 0.37 -0.23

81 7 0.19 -0.16 0.09 -0.38 0.40 -0.10 -0.04

68 7 0.42 -0.41 0.09 0.17 0.06 -0.14 -0.20

70 6 0.19 X 0.19 0.26 0 15 -0.68 -0.11

42 7 0.56 -0.27 -0.72 0.31 0.20 -0.00 -0.07

69 6 0.55 X -0.85 0.62 0.51 -1.07 0.24

24 5 0.13 0.34 0.45 -0.37 X X -0.56

67 7 0.42 0.23 -0.66 -0.09 -0.19 -0.14 0 43

40 7 0.26 0.78 0.39 0.51 -1.81 -0.72 0.59

56 5 -0.19 0.55 1.06 -0.66 -0.75 X X

54 5 0.39 0.42 X 0.06 -0.48 X -0.28

48 5 X -0.74 1.37 -1.25 -0.44 X 1.07
47 5 X -0.74 1.37 -1.25 -0.44 X I .07
55 5 -0.23 -0.18 0.54 -0.24 0.12 x X

53 5 -0.59 -0.24 X 0.61 0 07 Y 0.16

41 7 -0.93 -0.27 -0.75 -0.33 -0.86 1.60 1.54

34 7 -0.08 -0.02 -0.11 -0.08 0.28 -0.37 0.37

22 6 0.55 0.39 -0.09 -0.51 X -0.02 -0.32

8 7 1.27 0.90 0.01 -0.41 0.40 -0.54 -1.63

23 5 0.46 -0.59 0.52 -0.29 X x -0.10

39 7 -0.11 -0.16 -0.24 0.18 -0.34 0.74 -0.07

33 7 -0.35 0.31 -0.17 -0.20 -0.27 -003 0.71

21 6 -0.21 -0.05 -0.13 0.25 X 0.41 -0.27

7 7 -0.37 -0.31 0.01 0.39 1.20 -0 08 -0.84

this regards a lowest occurrence which would be situated too high, the
extended range chart is not affected by it. The modified RASC results had
 309

Fig. 8.13 Comparison of range charts obtained by five different methods for Palmer's database. Modified
RASC and RASC results were added to ranges previously plotted by Edwards (1982). Lowest and highest
occurrences were ranked for each method and these ranks were used to display the ranges. The modified
RASC, Shaw (1964) and Edwards (1978) results are similar. The Hay (1972) and RASC ranges were
based on average highest and lowest occurrences. These generally a r e shorter than the other
(conservative) ranges.

converged almost completely after 7 iterations as can be seen in Table 8.22

by comparing 326 to 327. The three sets of event variances (s25, s26 and s2 7)
for deviations from the spline-curves after five, six and seven iterations
are reasonably close. It is noted that the process of convergence shows
oscillations for some events k e . those for which the value of S26 is not
between those of s25 and s27 in Table 8.22).
The average deviation is clearly negative for the first six events in
Table 8.22. The same phenomenon was previously encountered for the
average deviation of the Gradstein-Thomas database (see Fig. 8.6) where
it was accompanied by a positive average deviation a t the bottom of the
stratigraphic sequence. The reason for this calibration problem can be
understood by comparing spline-curve and data in Figure 8.12 for the
Morgan Creek section. A decrease in the smoothing factor, which is
equivalent to increasing all event standard deviations by the same factor,
would result in a curve instead of a straight line for the Morgan Creek
section. This curve would be closer to the first five events in this section
310

(cf. Fig. 9.23 and later discussion in Section 9.10). It may be assumed that
the calibration problem of lack of fit near the tops and bottoms of some
sections is related to a slight overestimation of the smoothing factors
which, in turn, is equivalent t o a slight overestimation of the event
variances in these sections.
The sample sizes (n)of the events in the Palmer’s database are too
small to decide which events have variances that are significantly smaller
or larger than average. Neither can it be decided from the skewness
statistics in the last column of Table 8.22 which events have a n
asymmetrical frequency distribution. It is interesting that the five largest
(positive) skewness values (events 39, 33, 55, 41 and 7) are for lowest
occurrences whereas the two smallest (negative) skewness values (events
70 and 40) are for highest occurrences. All remaining events have
skewness values which are less than 0.90in absolute value.
The preceding observation would support the hypothesis t h a t
Palmer’s trilobites satisfy the model advocated by Edwards (see Fig. 2.13)
and Baumgartner (see Fig. 2.14). For the latter model, a lowest occurrence
has its longest tail pointing in the stratigraphically upward direction
(positive skewness) whereas a highest occurrence has its longest tail in the
stratigraphically downward direction (negative skewness).
 311

CHAPTER 9
EVENT-DEPTH CURVES AND MULTI-WELL COMPARISON

9.1 Introduction

This chapter describes the theory and application in geological basin

analysis of CASC for Correlation and Scaling in time. The CASC method
of quantitative correlation is based on the RASC method and on the
philosophical reviews and statistical methodology of several geologists and
mathematicians including: Shaw (19641, Hay (1972), Drooger (1974),
Blank (1979), Reinsch (19671, De Boor (1978) and Eubank (1988). The
method provides a precise, automated and semi-objective means of
correlation of rock sections for which an optimum sequence or a scaled
optimum sequence of biostratigraphic events has been determined using
the zonation method and computer program RASC. The next two sections
on principles of correlation and scaling in time and generalized description
of CASC method consist of material only slightly modified from
Gradstein et al. (1985), Agterberg et al. (1985), and Agterberg and
Gradstein (1985). This introductory material is followed by explanations
of the cross-validation and jackknife methods. Next this chapter contains
a number of regional applications of CASC. These include a comparison of
CASC with other methods of biostratigraphic correlation using Palmer’s
database for example. Comparisons between automated stratigraphic and
more subjective (manual) correlation are given for the Mesozoic and
Cenozoic microfossil record of the NW Atlantic margin. Other topics to be
discussed are integration of biostratigraphic and lithostratigraphic
information for the Central and Viking Grabens, North Sea, and
integrated CASC correlation of foraminifera1 and dinoflagellate datasets
for the Labrador Shelf and Grand Banks.
The use of RASC and CASC provides the stratigrapher with an
integrated biostratigraphic method, particularly suitable for exploiting
the considerable amount of micropaleontological data that accumulates
during sedimentary basin analysis. The method starts with a data file of
the original observations on the distribution in time and in space in wells
or outcrop sections of all taxa identified. Next, this data file is reduced t o
biozonations that best explain the regional and temporal trends. Finally,
312

geologically reasonable correlations of the sections will be calculated.

Segmentation and correlation of the original sections can be achieved by
means of fossil events and RASC biozones. Interpolation of the scaled
optimum sequence in linear time makes it feasible t o correlate by means of
isochrons. Each correlation line carries an uncertainty limit, which is a
combined estimate of various original uncertainties in the data.

9.2 Principles of correlation and scaling in time and comparison

with composite standard method
As previously discussed in Chapters 1 and 2, geological correlation
traditionally is expressed in terms o f (1) rock type units such a s
formations or well log intervals (lithostratigraphic correlation), (2) fossil
units such as biozones (biostratigraphic correlation), (3) relative age units
or stages (chronostratigraphic correlation) and (4)linear time units or
ages (geochronologic correlation). Instead of using units with a certain
thickness or a duration in time, correlation frequently is based on events.
Events or datum planes refer to fossilized, physical or organic occurrences
of supposedly irreducible resolution along the geological time scale.
An important contention of geological correlation is that once the
events or various types of rock, or relative age units have been properly
determined and defined, these units can indeed be used for correlation. As
pointed out by F.M. Gradstein in the Foreword, existing stratigraphic
codes show how to define stratigraphic units but they do not define how t o
correlate them. The actual correlation generally takes place in the
subjective domain of experts. Procedures for correlation or stratigraphic
equivalence depend on subjective evaluation of the unique relationship of
each individual record to the derived and accepted standard. It follows
that correlation as practised in geology cannot be readily verified without
a detailed review of all the underlying facts.

Traditionally there is no method of formulating the uncertainty in

fixation of individual records t o the standard. As Riedel(1979) stated:
“Biostratigraphy will be continued to be regarded as an art rather than a
science, until it is possible t o attach confidence limits t o suggested
correlations”. An improvement in definition of the zonation through
increased numbers of observations and taxa may increase the number of
correlation tiepoints, but still leaves the question of uncertainty
unanswered. Such an uncertainty generally is couched in qualitative
 313

terms only. In many geological investigations such a subjective procedure

yields satisfactory results, correlation being only a part of the scientific
objectives. Situations do arise, however, where the quality of correlations
determines the outcome of the study. This is particularly true in the field
of operational biostratigraphy, where large and complex data sets may
have t o be reduced before they can be of assistance in deciphering basin
history.
The problem of using subjective judgement only is not so much that it
leads t o right or wrong stratigraphy, but that a single solution is proposed.
It should be attempted t o establish reasonable criteria for successful
correlation by providing insight into the actual uncertainty in correlation,
either in millions of years or in depth in meters. In regional correlations
there frequently is limited or no understanding of how much (in depth or
in relative time units) the solution differs from alternate solutions, using
the same data. In all likelihood it is difficult to propose or compare two
alternative correlations, without major review or analysis of all
underlying facts.

Biostratigraphic correlation depends on the probability that: (a) in

each rock section the events defining a biostratigraphic increment have
been detected and properly taxonomically determined; and (b) the true (or
natural) sequence of events is known. This principle was succinctly stated
by Hay(19721, who then went on t o propose the principle of matrix
permutation for construction of the most likely sequence of (nannofossil)
events in time (see Chapter 5). In the ranked sequence each event position
is an average of all the relative positions but no direct insight is available
into the uncertainty of rank.
As early as 1964, Shaw not only proposed a simple ranking method for
biostratigraphic events, but also a method for correlating the sections in
which the events occur. The original method is as follows. From a number
of individual geological sections (A, B, C, D, etc.), one (for example A) is
selected that shows a relatively complete and reasonable “normal” order
and spacing of events. This particular sequence of entries and exits of taxa
is plotted along one axis and that of a comparable sequence B along the
other axis of a conventional two-axis graph. Scale units are in feet or
meters, as found in each section but, in a simplified procedure, order only
can be used. The best fit of the resulting scattergram is called the line of
correlation.
314

Shaw (1964) advocated regression analysis as a linear trend-fitting

technique although A and B probably are both subject to uncertainty. (By
subjectively deleting the larger deviations, Shaw avoided systematic
discrepancies due t o neglecting uncertainty in A as previously discussed in
Section 8.4). The order and spacing of first and last occurrence events
along the A-axis is now updated through projection of the homologous
B-axis events, through the best-fit line, onto the A-axis. If the first
occurrence of an event in B occurs relatively lower than in A, the range of
this event in A is extended downward. If a last occurrence of an event in B
occurs relatively higher than in A, the range of the event is extended
upward. It is attempted to maximize the stratigraphic ranges. Next the
updated A-axis (composite section) is compared to section C, in the same
manner as A was compared to B and the process is continued by including
an increasingly larger number of sections. In the final composite section,
the scale of the successive events has become a composite of all spacing
values between successive events. Because the final result depends on the
order in which individual sections were added to the composite section,
there may be a second or even a third round during which A, B, C, ... again
are plotted against the composite section.

Actual correlation of events is achieved by making new bivariate

plots for each individual section as a function of the final composite
standard. For each bivariate scattergram a new best fit line or best fit
channel is calculated which serves to project the composite events onto the
individual section scale. In a mathematical sense, each value in the
composite standard can be expressed as a function of its correlative (depth)
value in the individual sections. Miller (1977) provides a good description
of use of the composite standard method.

The CASC method of quantitative correlation combines average

sequence methodology with bivariate correlation technique. Input for
CASC is the RASC input file that shows the original sequence of events in
each of the sections. In addition, the program requires a depth file, that
shows the observed depth in feet or in meters for all the events in the
original sequence file. The correlation and scaling in time (CASC)
program first computes the RASC optimum sequence and RASC scaled
optimum sequence of events. Using the three normality testing
techniques in the RASC method (bivariate graphs, stepmodel and
normality testing), outliers in the individual sections may be eliminated.
Based on the filtered data file, a new optimum sequence may be calculated,
after which each individual sequence of events is compared to the scaled
 315

Adolphus D-50

line of observalion
(events versus depths)

I
I
1

\
3 16

optimum sequence, and best fit curves (smoothing splines, see later) are
calculated. A spline fit yields a function such that, for each optimum
sequence position, the most likely stratigraphic equivalent position can be
found in the individual sequences. These normalized tiepoints then are
correlated.
Figure 9.1 graphically depicts the principal steps, executed for the
correlation of event 29 (top of Cyclammina a m p l e c t e n s ) i n t h e
Adolphus D-50 well on the Grand Banks, which is part of the Gradstein-
Thomas database. The y-axis is the optimum sequence in 21 of the wells
( h , = 7, m,l = 2; probabilistic ranking followed by modified Hay method).
Instead of the optimum sequence, the scaled optimum sequence can be
used (see later). The x-axis is the observed sequence of events, whereas the
z-axis is the common depth scale of the well. The lower scattergram
expresses mismatch of the individual sequence and the optimum sequence.
The best fit line for the graph (here visually estimated) is the line of
correlation.
Working with event scales initially has the advantage that complica-
tions due t o different rates of sedimentation in different places which may
be hundreds of km apart are avoided. Moreover, equal spacing of values
for the independent (x-axis) variable in spline-curve fitting has the
considerable advantage that the possibility of unrealistic oscillations of
the fitted curve between irregularly spaced control points is avoided.
However, the number of levels in the event scale differs from section to
section in a “random” manner. For correlation between wells it is
necessary to replace the levels of the event scales by depths (in km). This
replacement is shown in the upper part of the scattergram of Figure 9.1.

The individual sequencex is a function of the depthz at which the

events were observed. This function is shown as the “line of observation”.
The most likely position of event 29 in the Adolphus well is found by
projecting its optimum position via the line of correlation t o the individual
sequence and from there via the line of observation to the depth scale.
Thus all optimum sequence events are scaled in (well) depth. In a multi-
well comparison, the most likely depth value (z-axis) in each well is
calculated for selected y-axis values (event positions) in the optimum
sequence. In the example, event 29 in Adolphus should occur at 6050ft
(observed 6200ft). In another well (Flying Foam, not shown here) event 29
was projected to occur at 4850ft (observed 5330ft). These depths then are
 3 17

the most likely correlation tiepoints and a line can be drawn t o connect
them.

The standard deviation (SD) of the events relative t o the line of

correlation (and the line of observation) in the y-direction (and parallel z-
direction, which is the depth-axis in the well) provides an estimate of the
mismatch of each event. Later on this will be called the local error. When
it is geologically unreasonable to expect a continuous sedimentation rate
in the vicinity of a certain depth, the local SD can be modified to account
for changes in the sedimentation rate.

The same procedures as shown in Figure 9.1, using the RASC

optimum sequence, can be applied when the scaled optimum sequence is
chosen instead. The interfossil distances in the scaled optimum sequence
reflect the average relative distance of the events in relative time. If it is
possible to estimate the numerical geological ages of some of the events in
the scaled optimum sequence, the relative distance estimates can be used
to stretch the scaled sequence in linear time. This way the scaled optimum
sequence becomes a (local) biochronology and hence isochrons can be
traced through the wells or outcrop sections. For paleontologists this is a
valuable method for finding the numerical age of the most likely position
of principal zone boundaries in each well. Such boundaries, as argued in
Gradstein et al. (1985, Chapter 11.4), can delineate sedimentary cycles.
The original standard deviations for the interfossil distances in the scaled
optimum sequence now reflect the uncertainty in linear time between the
events. This uncertainty can be expressed by means of the global error bar
(see later).

As a first test of the validity and use of this numerical time

interpolation for geological analysis, Figures 9.2 and 9.3 were constructed
by Gradstein et al. (1985). Along the vertical axis of Figure9.2 the
interfossil distances are plotted for the Cenozoic foraminifera1 events in a
scaled optimum sequence for 2 1 wells. For some events, listed later in
Section 9.8 on regional applications , which are the (regionally averaged)
last occurrences of key planktonic and a few benthic Foraminifera,
numerical ages can be estimated. This involves comparison of the regional
t o standard Atlantic zonations, details of which follow later. The
horizontal scale is linear time in Ma, for the Cenozoic period. In principle,
the numerical ages are for highest occurrences and not for average highest
occurrences (see Fig. 8.9). On average, there is about 10 m.y. difference
between these two types of stratigraphic events in the Atlantic zonations
318

considered. A systematic discrepancy of this type, however, is eliminated

automatically provided that is approximately the same for the different
taxa used.

60 50 40 30 20 10 0
I

ri
2i:q t
1

RASC interfossil distances

____)
Ma

228

16

I
184

50
46
54

56

55

59
61

RASC TIMESCALE

Fig. 9.2 Plot of the Cenozoic scaled optimum sequence (21 wells; 7/2/4 run) versus linear time in Ma. The
inter-event distances are plotted cumulatively. For some selected events in the scaled optimum sequence
the numerical age is known (dots),and this allows to scale the whole fossil sequence in linear time.
 319

The calibrated events are used to form a nomogram for correlation, so

that all events in the scaled optimum sequence can be dated. In Figure 9.3

Biochronology

60 50 40 30 20 10 Ma o
Pal. Eocene Oligocene Miocene
1000 -

2000.

Adolphus D-50

3000 -

r
4000-
.-
1
C

5
a
2
5000.

6000 -

7000 -

@ subjective age
8000.
RASCage

9000 -

SEDIMENT ACCUMULATION

Fig. 9.3 The RASC biochronology of Fig. 9.2 is used to estimate rate of sedimentation (dashed line) in the
Adolphus well. The solid line (subjective) shows approximately the same trend, using independent well
history data (from Gradstein and Agterberg, 1985).
320

this new RASC biochronology (horizontal axis) is used to estimate the rate
of sediment accumulation (dashed line) in the Adolphus D-50 well.
Several years earlier, prior t o development of RASC and CASC a n
approximate chronostratigraphy of this well section had been given, in
system units from the Paleocene upward. As shown in Figure 9.3, there is
a close approximation of the two, independently arrived at, sediment
accumulations. The earlier interpretation obscured a possible late
Oligocene-early Miocene hiatus. Scaling in time of the scaled optimum
sequence is a practical way of erecting a regional time scale.

In summary, the CASC method of correlation is based on three

conditions: (1) each individual stratigraphic sequence of events is a
sample of the optimum sequence; (2) the observed depths of the events in a
stratigraphic section are estimates of the true depths; and ( 3 ) the
calculated relative interfossil distances of events in the scaled optimum
sequence can be used to stretch this sequence along the numerical
geological scale with known ages for index fossils in this sequence
providing the necessary tiepoints. Input for automated correlation of fossil
events (or zones) by means of CASC with confidence limits in depth or time
units are: (a) depth in feet or in meters for all fossil events in all wells or
outcrop sections. These events are the same as those used in the RASC
method; (b) ages of index fossils to stretch the scaled optimum sequence in
linear time; (c) events, clusters of events (zones) or ages to be correlated;
and (d) wells or outcrop sections to be correlated.

9.3 Generalized description of the CASC method

Originally, the CASC program (Agterberg et al., 1985)was developed
on a CDC Cyber 730 mainframe computer with a Tektronix 4014 terminal
with code in FORTRAN Extended Version 4. Two computer libraries were
required t o use CASC: IMSL Library and Tektronix Advanced Graphing
Library. Also, mass storage facilities were used. It was assumed that in
order to obtain the geologically most satisfactory bivariate fits, mainframe
CASC was best used interactively.

One of two different routes could be selected at the beginning of an

interactive CASC session. The first route uses as a starting point the
RASC optimum sequence, plotted against the so-called event scale. The
latter has entries for the original sequence data in each stratigraphic
section. Instead of the optimum sequence, the RASC scaled optimum
sequence may be used. The latter combines average order and relative
distance. The optimum sequence option is simpler than the so-called
distance option based on the scaled optimum sequence, but not principally
different. As an additional illustration the distance method will be applied
t o RASC results for the distribution of Cenozoic foraminifers in
offshore wells on the Grand Banks and Labrador Shelf.

If RASC distances were used as input in mainframe CASC, it was

required to replace them by ages (in Ma). This replacement is not required
in the CASC modules of micro-RASC (see Chapter 10). The procedure in
mainframe CASC is schematically shown in Figure 9.4. Assuming
approximate ages are known for a subgroup of events in the scaled
optimum sequence shown in Figure 9.4, the objective is to fit a curve t o
these data in order to be able to replace any RASC distance by its age (see
Fig. 9.4d). First, RASC distances with the same age are averaged (see Fig.
9.4a). Then a cubic spline curve is fitted t o the age-distance pairs

++-o
0 B
0

0 0

RASC distance RASC distance

GSC

Fig. 9.4 Schematic illustration of method followed in CASC mainframe computer program to establish
relation between RASC distance and age. (a) Two (or more) RASC distances for the same age a r e
averaged. (b) Cubic spline-curve is fitted using age as the dependent variable; smoothing factor (SF)
representing standard deviation of differences between event ages and curve is chosen in advance, before
curve-fitting. (c) Standard deviation (SD)for differences between original values and curve is computed
after curve fitting. (d) Fitted curve is used to convert any RASC distance into corresponding age.
322
minimizing the sum of squares of deviations between points and curve in
the vertical (age-) direction of Figure 9.4b. The smoothing factor SF can be
chosen beforehand by the user of the interactive CASC computer program.
It is equal t o the square root of the mean square deviation between points
and curve. Because this standard error generally is not known
beforehand, the user can determine i t by trial and error while
experimenting with different plots on the screen of the monitor. In Figure
9.4b a curve was fitted to 5 original values (0)and 2 averages of two values
+
( ). The standard deviation of the original data in relation to this curve is
also shown on the screen (SD in Fig. 9.4~). The fitted curve does not
extrapolate outside the range of the RASC distances used for the curve
fitting. Consequently, the circle with the highest RASC value is not
considered for estimating SD in this example. It is noted that a curve also
could be fitted directly through the 8 circles in Figures 9.4a and 9 . 4 ~ Then
.
SD would be equal to SF.

RASC distance RASC distance

< 0 < 0

(a) 0 -1 (C)

0-+4
0
0
?+4
0
0

Fig. 9.5 Schematic illustration of preliminary computing and optional editing procedure a t beginning of
CASC mainframe computer program. (a) Events found to be anomalous with a probability of over 99 per
cent (asterisk) may be omitted from spline-curve fitting and later plots; RASC distances of two (or more)
coeval events are averaged. (b) Cubic spline-curve is fitted using RASC distance as the dependent
variable; smoothing factor (SF) representing standard deviation of differences between RASC distances
assigned to levels and curve is chosen in advance. (c) Standard deviation (SD)is computed from
differences between original values and curve after curve-fitting; original values (e.g. those labelled R)
can be deleted. (d) New curve with new standard deviation (SD)is obtained without use of deleted
values.
 323

Next, the CASC user can display and edit the RASC distances for any
well from the set of the wells used. Editing options are schematically
shown in Figure 9.5, which displays preliminary data analysis. The scale
in the vertical direction is relative. It shows successive levels for the
stratigraphic events in the well considered. RASC distances of 2 or more
coeval events are averaged (see Fig. 9.5a) before cubic spline fitting (Fig.
9.5b). The user has the option of omitting events for which the second-
order differences were anomalously high (i.e. shown by two asterisks in
the RASC normality test). Such anomalous events are then displayed by
use of a special symbol (single asterisk in Fig. 9.5a) and are not employed
for curve fitting. The deviations are measured in the horizontal direction.
SF and SD serve the same purpose as in Figure 9.4. The user may wish t o
remove other events considered t o be anomalous, for example, those
labelled R in Figure 9.52. Then a new cubic spline-curve will be fitted for
the reduced data set (Fig. 9.5d). If extreme values are deleted, SD will
probably be decreased in value. The original RASC model is based on the
assumption that positions of events in a well are distributed around their
expected value, according t o a normal probability distribution, with
standard deviation set equal to l N 2 = 0.7071. One, therefore, would
expect SD t o be approximately equal to 0.7 if the number of events in the
stratigraphic section is sufficiently large.
For further analysis in preparation of automated correlation, RASC
distance is replaced by age (see Fig. 9.6a) using the earlier derived
relationship between RASC distance and age (see Fig. 9.4d). In the
following discussion, the variables for event level, age and depth are
denoted as x, y and z, respectively. A spline-curve can be fitted to express y
as a function of x, as was done for distance in Figure 9.5d. It also is
possible to replace the levels by their depths and fit a spline curve t o
express y as a function of z using depth as the independent variable. This
leads directly to a plot similar t o Figure 9.6f.
However, the rate of sedimentation may have changed significantly
during geologic time at a well site and this can result in irregular
distribution of the points along the z-axis. This, in turn, may make it
difficult t o obtain a spline-curve that extrapolates in a satisfactory manner
across data gaps along the z-axis corresponding t o short periods with
increased sedimentation rates (also see Section 3.6). For this reason, the
indirect method given in Figure 9.6 can be employed instead. Assume that
the spline-curve of Figure 9.6a is written a s y = f(g) + ey where ey
represents a random deviation in the y-direction. The bar under x
324
indicates that y is regressed on x using data points which are regularly
spaced along the x-axis. Depth ( z ) is plotted against x in Figure 9.5b,and a
+
separate spline-curve with z = g(3) e, is obtained, using the same set of
regularly spaced data points along the x-axis. The deviation e, points in

15 10 5
v< 2p I yo.0
- 0.2

- 0.4
- 0.8
- 0.8
- 1.0
- 1.2
I I I I I I 1.4
0.0
1 2 3 4 5 6
Level ( x - )

02 (0 -02

Jl-;
04 -04

O6 9 - hfzJ
-08

-08
O8 ii
10 8 10

12 12

14 14
V V

Fig. 9.6 Schematic illustration of calculation of an event-depth curve from RASC output for a well.
(a)RASC distances have been replaced by ages using relation illustrated in Fig. 9.4d; new spline-curve f
(xJ is fitted; bar in x_ denotes use of regular sampling interval for x ; smoothing factor (SF), which was
selected before curve-fitting using one age per level, is smaller than standard deviation (SD)for all
original values. (b) Spline curve g (2)is fitted to express depth as a function of level x ; bar in Zdenotes use
of regular sampling interval for 2 ; SF= SD is equal to some small value. (c) P represents spline-curve g (z)
in Fig. 9.6b now coded as set of values for x a t regular interval of z. (d) Q, denotes curve passing through
set of values of y a t regular interval of z obtained by combining spline-curve of Fig. 9.6a with that of
. 9 is spline-curve fitted to values yzx of Fig. 9.6d using new smoothing factor SF. (0 Standard
Fig. 9 . 6 ~ (e)
deviation SD is computed after curve-fitting, using one age per level.
 325

the z-direction. Obviously, the curve g(z) cannot decrease in the x -

direction.

The curve for z in Figure 9.6b again is shown in Figure 9 . 6 ~ .It has
been rewritten in the form 32 = g-'(g), t o indicate that estimates f were
obtained at points which are regularly spaced along the z-axis. Assume
that j is obtained for the irregularly spaced values of x in Figure 9 . 6 using
~
f i x ) shown in Figure 9.6a. This results in a set of values of j , , = fig-'(,))
for regularly spaced points along the z-axis (see Fig. 9.6d). The function
fig-1(g)) is not a simple mathematical expression. For example, its first
derivative is not readily available. A cubic spline j = h(z_)can be fitted to
the values j , , (see Fig. 9.6e). In Figure 9.6e, j is considerably smoother
than j X z .By using a smaller smoothing factor (SF), the difference between
j and j x zmay be kept negligibly small (see curve to be used for example in
Fig. 9.7a). The standard deviation SD for points used for fitting in Figure
9.6a with respect to the curve 4 is provided in Figure 9.6f. The deviations
from j are measured in the y-direction. A similar age-depth diagram is
shown in Figure 9.7a where less smoothing was applied. The spline-curve
j = h(z)can be used t o assign a probable age t o any point along the well.

Figure 1.2 in Chapter 1 showed a so-called multi-well comparison for

five wells. It is based on a 7/2/4 RASC run on 21 wells. Points with
estimated ages of 10, 30, and 50 Ma along the five wells are connected by
lines of correlation in Figure 1.2. The uncertainty in the position of these
isochron contours is indicated by error bars, constructed according to one
of three methods further explained in Figures 9.7 and 9.8. The displays of
Figures 1.2a and 1 . 2 ~were redrafted from displays on the Tektronix
terminal obtained during an interactive CASC session; the error bars in
Figure 1.2b were obtained from event-depth plots according t o the method
explained in Figure 9.7d.

The local error bar in Figure 1.2a was obtained by multiplying s(y)
( = SD) along the y-axis by rate of sedimentation to obtain a modified error
s(z) along the z-axis, as shown in Figure 9.7b. The rate of sedimentation
(Fig. 9 . 7 ~is) the first derivative dzldy for z in j = h ( z ) . In general, a cubic
spline curve y, fitted t o n data points, consists of (n-1) successive cubic
polynomials
326

I I I I I I ,
- 0.2
(a)
- 0.4
- 0.6
- 0.8 5

1.o

1.2

I I l l I I

I J.
GSC

Fig. 9.7 Schematic illustration of estimation of local error bar and modified local error bar. (a) Standard
deviation SD was computed after curve-fitting, using one age per level. (b) Error bar of age value plus or
minus SD along Y-axis is transformed into error bar along Z-axis using first derivative (dzldyl of age-
depth curve. (c) Rate of sedimentation (=dz/dy) can be displayed on screen during CASC interactive
session. (d) Modified local error bar is asymmetrical with respect to depth value for a given age.

/
0

- I
RASC distance
2u
I

GSC

Fig, 9.8 Schematic illustration of estimation of global error bar. Theoretical standard deviation
a (=0.7071) along RASC distance scale is assumed to remain constant. It is transformed into variable
SD along age scale (e.g. SD'and SD").
 327
y = y, t cl,d + c2,d2 t c3,d3
(9.1)

with d = z-zi,zi 5 z < zi+l, where zi and z i + l (i = 1, 2, ..., rz-1) represent

the n depths used to convert j x = f i x ) into 9 = h k ) . The coefficients cli, cgi
and c3i can be used to calculate

d y l d z = cl, + 2c2,d + 3c3,d2 (9.2)

at any point. Inversion of this expression gives dzldy. The new standard
error s(z) = (dz/dy) s(y) can be displayed for any z as the local error bar
z k s(z) (see Fig. 1.2a). This propagation of error is based on the local rate
of sedimentation, which is assumed t o remain approximately constant
over the interval y +_ s(y). The latter condition frequently is not satisfied,
especially when j has many inflection points (between local maxima and
minima in sedimentation rate). Curvature of j t is considered in the
construction of a modified local error bar as illustrated in Figure 1.2d. For
any point z = h-l(y), this bar extends from the point h - l b - s(y)} t o
h - l b + s(y)}. It is asymmetrical with respect to z and is significantly
shorter at places where the rate of sedimentation is high.
Finally, a global error bar (Fig. 1 . 2 ~can
) be constructed as illustrated
in Figure 9.8. The standard deviation u = l l d 2 of events along the RASC
linear scale for distance is changed into a variable standard error s(y)
along the age scale. This new, variable standard error is changed into s(z)
according t o the method used for SD in Figure 9.7b. In global error bar
estimation, it is assumed that a single RASC distance error u can be
applied to all wells. On the other hand, in local error bar estimation, use is
made of a constant SD value along the age scale which was estimated from
the deviations between the points used for spline-fitting and the spline
curve itself (cf. Fig. 9.6e). Because of possible elimination of anomalous
events and averaging of ages for events at the same levels, the local error
bar is likely to be narrower than the global error bar. It is possible that the
quality of the biostratigraphic information is not the same in all sections
considered. Such differences would be considered in local error bars but
not in global error bars.
The purpose of the error bar is to quantify the uncertainty of the
observed depths of events with respect t o their estimated depths in the
wells. Each local or global error bar extends from the estimated depth
328

minus one standard deviation t o estimated depth plus one standard

deviation. If the observed depth is normally distributed about the
estimated depth, the error bar can be interpreted a s a 68percent
confidence interval for single events. Then there is a 68percent
probability that the observed event falls within the range outlined by the
error bar. Likewise, there is 95 percent probability that the observed event
falls within an extended error bar, which is 1.96 times as wide as the error
bar shown.
It will be shown in Section 9.7 that the actual precision of the
estimated depth of an event in a well can be greater than that indicated by
the error bar for single events. If a value on the spline-curve would be
interpreted as the arithmetic average of all (n) values used for its
estimation in a well, the standard deviation of this mean would be equal to
the standard deviation of the single events, divided by d n . For example,
in a well with 16event levels, the standard deviation of the difference
between estimated depth and “true” depth would be one-fourth the
standard deviation of the 16 single events used t o construct the error bar.
The degree of validity of the assumption that a value on the spline-curve
can be interpreted as an arithmetic mean of all values used for estimation
is not precisely known, except when the smoothing factor is large. In the
limit, which is reached when SF exceeds an upper threshold value, the
spline-curve reduces to the best fitting straight line of least squares. Then
the preceding assumption holds approximately true.

Output from a 7/2/4 RASC run on 2 1 wells and a 5/2/3 run on 24 wells
were used as input for examples of actual CASC runs in the remainder of
this section. Table 9.1 shows the optimum sequence, modified optimum
sequence (after final reordering) and RASC distances for the 7/2/4 run on
21 wells (also see Fig. 6.2). Several events, occurring in fewer than
sevenwells, were later inserted as unique events. Table 9.2 shows
estimated ages of 22 events, including these unique events. Average
RASC distances for events with the same age are shown in the last column
of Table 9.2. Figure 9.9a shows the ages plotted against the RASC
distances. The displays in Figure 9.9 (and Figs. 9.10-12) were redrafted
from hardcopy of displays on a Tektronix terminal. A cubic spline function
with smoothing factor SF = 2.0 was fitted to the 15 ages, using the
average distances shown in the last column of Table 9.2. The smoothing
factor SF is the standard deviation of differences between the 15 ages and
corresponding estimated ages on the spline-curve for the same RASC
 329

TABLE 9 . 1

RASC output for 7 / 2 / 4 run on 21 wells (Grands Banks - Labrador Shelfl used as CASC input. Event
levels (sequence position numbers 1-40) (A), optimum sequence of events identified by their dictionary
numbers (B), modified optimum sequence after final reordering (C), and cumulative RASC distances for
events in last column.

W C W C
A B C Distance A B C Distance

1 in 10 n. 000 21 261 261 4.364

2 17 17 0.391 22 263 263 4.518
3 16 16 0.912 23 40 40 4.645
4 67 67 1.204 24 29 32 4.802
5 21 18 1.647 25 32 29 4. 809
6 18 71 1.799 26 41 264 5.175
7 71 21 1.919 27 264 42 5.189
a 20 20 2 . 108 28 42 41 5.237
9 26 15 2.442 29 30 30 5.251
10 15 26 2.486 3r) 86 90 5.531
11 70 70 2.513 31 36 36 5.620
12 27 29 3.198 32 90 86 5.667
13 69 25 3.358 33 45 45 5.786
I4 24 27 3.418 34 57 57 5.799
15 25 69 3.499 35 50 50 6.302
16 81 81 3.722 36 46 46 6.429
17 259 259 3.738 37 54 54 6.738
18 33 33 4.295 38 56 56 7.178
19 34 34 4.31 I 39 55 55 7.689
20 260 260 4.342 40 59 59 8.033
I

TABLE 9.2

Estimated ages for 2 2 events and calculation of average RASC distances for two or three events with
same estimated age.

Event Age RASC Average Event Age RASC Average

No. (Ma) Distance Distance No. (Ma) Distance Distance

4 3.5 -0.58s -0.385 85 38 4.456 4.456

5 3.5 -0.476 29 40 4.509 4.809
269 3.5 -0.096 90 49 5.531 5.531
17 11 0.391 0.391 57 52 5.799 5.247
179 15 1.984 1.984 93 52 5.895
15 17 2.442 2.442 50 55 6.302 6.302
26 20 2.486 2.427 194 57 7.073 7.073
137 20 2.368 55 58 7.689 7.434
24 28 3.198 3.198 56 58 7.178
33 37 4.295 4.017 61 63 8.228 8.039
259 37 3.738 253 63 7.849

I
330

distance values. The standard deviation of the 22 original ages before

averaging of some RASC distances is also shown in Figure 9.9a.
The original sequence of events occurring in 7 or more wells is shown
in Table 9.3 for Indian Harbour M-52(Well No. 5), which will be used for

a AGE IN M a

7 0

s - a m
m-z.sim

:t
C AGE IN Ma d FIRST D E R I V A T I V E
7 0 e O M 4 0 9 0 2 0 1 0 0 10 9 8 7 3 6 4 3 ? ? 0

m
<
m
z
-10 -I

e EVE N T LEVEL f AGE IN Ma

30 26 20 16 10 6 1
>

0
m
-I
0

2 =
-
z
x
c
3

4 L4

Fig. 9.9 Example of CASC displays for Indian Harbour well based on 7/2/4 RASC results for 21 wells. ( a )
Age-RASC distance relationship a s derived from the 21 wells file. (b) Initial CASC plot for default
smoothing factor. ( c ) Age-level plot for default SF. (d) First derivative of ( c ) . (e) Level-depth plot. (fl
Age-depth plot for default SF; spline-curve was fitted directly to the data, using irregularly spaced
depths.
further analysis. As mentioned before, one of two different routes can be
selected at the beginning of mainframe CASC. These consist of using
either optimum sequence data or RASC distances for the events. In both
subprograms, event levels for successive, non-coeval events are defined, as
illustrated for Indian Harbour M-52, in the second column of Table 9.3. In
the second subprogram, the RASC distances in a well are transformed into
ages using the spline-curve fitted in Figure 9.9a (see last column in Table
9.3). The methods used in the two subprograms are identical, except that
sequence position numbers instead of ages a r e used in the first
subprogram. Only the option that uses the ages (in Ma) will be illustrated
in detail here.
Mainframe CASC produces a number of successive plots. For each of
these plots the user is required to answer one or more questions. The plot
that comes after Figure 9.9a during a CASC session is shown in Figure
9.9b. It shows the RASC distances of Table 9.3 plotted against their event
levels. Before this plot is actually shown on the Tektronix screen, the user
is asked if he wishes to exercise the option of deleting anomalous events
which are out of place with a probability of 99percent according t o the
RASC normality test. Moreover, points can be deleted from Figure 9.10

TABLE 9.3

CASC input for Indian Harbour well; definition of 18 event levels; and transformation of RASC distances
into ages using spline-curve in Fig. 9.9a.

Event Event RASC Age Event Event BhsC Age

No. Level Distance (Ha) No. Level Distance (Ha)

10 1 0.000 5.6 34 12 4.311 37.0

18 2 1.647 15.2 263 13 4.518 38.9
15 3 2.442 21 .o -36 13 5.620 47.9
-20 3 2.108 18.6 29 14 4.809 41.4
-16 3 0.912 10.9 -40 14 4.645 39.9
17 4 0.391 7.9 -4 1 14 5.237 44.2
24 5 3.198 27.3 -42 14 5.189 44.4
-25 5 3.358 28.7 86 15 5.667 48.2
26 6 2.486 21.4 45 16 5.786 49.1
-27 6 3.418 29.2 -46 16 6.429 53.7
259 7 3.738 32.0 57 17 5.779 49.2
261 8 4.364 37.5 -54 17 6.738 55.6
30 9 5.251 44.9 -50 17 6.302 52.9
260 10 4.342 37.3 55 18 7.689 61.5
-32 10 4.302 41.2 -56 18 7.178 58.4
33 11 4.295 36.9 (59)
332
b SEDIMENTATION RATE

C AGE IN Ma d FIRST DERIVATIVE

m w m 1 o 5 o z a i o o ID e II 7 a s 4 a o 1 o

. / *

S
m
m
C

10 ;
m
m
1s

f SEDIMENTATION RATE
1 0 0

Fig. 9.10 Example of CASC displays for Indian Harbour well (continued from Fig. 9.9). (a) Spline-curve
for small (default) SF fitted to combination of Figs. 9 . 9 ~
and 9.9e; indirect method explained in Fig. 9.6
was used. (b) Sedimentation rate in 0.1 k d m y (=first derivative of spline-curve in Fig. 9.10a multiplied
by 10); local maximum and minimum are due to lack of smoothness of spline-curve as explained in text.
(c) Age-level plot for SF=4.0 instead of default, used in Fig. 9 . 9 ~ .(d) First derivative for Fig. 9 . 1 0 ~ ;
magnitude of peak in Fig. 9.9d has been reduced. (e) Spline-curve for small (default) SF fitted to
combination of Figs. 9.9e and 9 . 1 0 ~ (0
. Sedimentation rate in 0.1 k d m y corresponding to Fig. 9.10e.

itself by positioning the Tektronix cursor on top of them. No points were

omitted in this example. Next a cubic spline-curve is fitted to the average
RASC distance values in the third column of Table 9.4. First the user is
shown the default smoothing factor (SF = 0.5146 for Fig. 9.9b) and asked
if this value should be used. This default was obtained automatically, by
fitting spline-curves with SF increasing from 0.0 until the first curve is
 333

TABLE9.4

Data used for fitting spline-curves in Indian Harbour well example shown in Figs. 9.9 to 9.11

Event Depth Average Averaie Event Depth Average Average

Level (n) Distame Age Level (m) Distance Age

1 546 0.000 5.6 10 1912 4.572 39.3

2 619 1.647 15.2 11 2045 4.295 36.9
3 720 1.821 16.8 12 2305 4.311 37.0
4 747 0.391 7.9 13 2335 5.069 43.4
5 1067 3.278 28.0 14 2366 4.970 42.6
6 1232 2.952 25.3 15 2396 5.667 48.2
7 1616 3.738 32.0 16 267 1 6.107 51.4
8 1674 4.364 37.5 17 2884 6.280 52.6
9 1732 5.251 44.9 18 3000 7.434 59.9

found for which the distance does not anywhere decrease with increasing
depth. The default solution is shown in Figure 9.9b. The smoothing factor
is the standard deviation of the differences between the 18 average RASC
distances and the fitted spline curve. The standard deviation of residuals
( = 0.5664) representing differences between original RASC distances and
fitted spline-curve is also given in Figure 9.9b. It is noted that this value is
only slightly less than u = 0.7 representing the theoretical standard
deviation along the RASC scale.

Figure 9 . 9 ~shows a new default result, obtained after replacing

RASC distance by age. It is possible to inspect the first derivative dxldy of
this graph (Fig. 9.9d). If the slope in the direction of increasing age for the
exceeds 10, its values are not displayed in Figure 9.9d.
curve in Figure 9 . 9 ~
Because the default yields the first monotonically increasing spline-curve,
normally a t least one interval with very high sedimentation rate is
introduced with this option. By increasing SF, the user can remove
artificially high sedimentation rates.

Figure 9.9e shows the relationship between depth and event level
with fitted spline-curve for SF = 0.02. It passes almost exactly through
the observed values. After display of this plot, the CASC user has the
option of either using this spline-curve in conjunction with the age-event
level plot of Figure 9.9b, or of by-passing the indirect procedure by directly
fitting a curve to the event-depth diagram in which event levels have been
replaced by their depths. The default result for the direct method is shown
in Figure 9.9f.
334

The result obtained by following the indirect method is shown in

Figure 9.10a for small SF (=0.1). The first derivative corresponding to
Figure 9.10a is given in Figure 9.10b. The irregularity between 2.2 and
a AGE I N Ma b 1 AGE I N Ma
70 W W 40 SO 10 10 0 70 80 50 40 30 20 10 0
t

d EVENT LEVEL
30 25 20 15 10 5 1
0

/
1
0
m
+
P
I
Z Z
X
<

f SEOIMENTATION R A T E

1
0
V
I
2 2
X
<

Fig. 9.11 Example of CASC displays for Indian Harbour well (continued from Figs. 9.9 and 9.10. (a)
Unsmoothed combination of Figs. 9.9e and 9.10~;note similarity with spline-curve i n Fig. 9.9e for
SF=O.l. (b) Curve of Fig. 9.11a smoothed with SF=O.l. (c) Sedimentation rate in 0.1 kmlmy
corresponding to Fig. 9.11b. (d) Level-depth plot for SF=0.0. (el Spline-curve for small (default) SF
fitted to combination of Figs. 9 . 1 0 ~and 9.11d; note similarity with spline-curve in Fig. 9.10e. (fl
Sedimentation rate in 0.1 k d m y corresponding to Fig. 9.11e; local maxima and minima are due to lack
of smoothness of spline-curve as explained in text.
 335

2.3 km in this diagram is due to lack of precision of the approximations,

used in the indirect method, to obtain new values on the spline curve (see
Figs. 9 . 6 ~and 9.6d). The regular spacing along the depth scale, used for
this purpose in mainframe CASC, is 50 m. Consequently, irregularities
due to lack of precision will not extend for more than 100m along the
depth scale.
Figures 9 . 1 0 ~t o 9.lOf show new results for the indirect method,
obtained after changing the value of SF from the default (SF = 3.58 in Fig.
9 . 9 ~ t)o SF = 4.0 in Figure 9.10~. During a CASC session, the user is
shown the unsmoothed values of f(g-'(Z)} (cf. Fig. 9.6d) connected by
straight lines. An example of the latter type of display is Figure 9.11a
which originally appeared during the CASC session just before Figure
9.10e where SF = 0.1. It is not possible to see differences between the
curves of Figures 9.10e and 9.11a. However, when SF is enlarged t o 1.0,
the smoother curve in Figure 9.11b is generated from Figure 9.11a. The
rate of sedimentation for Figure 9.11b is shown in Figure 9.11~.
Figure 9.11d represents the depth versus event level curve that
replaces Figure 9.9e, when SF = 0.00 instead of SF = 0.02 is selected. The
difference between these two curves is small and when the curve of Figure
9.11d is combined with that of Figure 9.1Oc, the resulting plot (Fig. 9.11e)
does not differ significantly from Figure 9.10e. However, the first
derivative of Figure 9.11e which is shown in Figure 9.11f differs
significantly from Figure 9.10f. It shows many 50 m irregularities, which
are due t o lack of precision as also discussed before (see Fig. 9.10b). In
general, the final event-depth curve is less sensitive t h a n t h e
sedimentation rate curve to small changes in the choice of smoothing
factors for successive curves during an interactive CASC session.
As a final example, Figure 9.12 contains various CASC displays for
Adolphus D-50. Table 9.5 shows the input for this CASC run which
consists of the partial DAT file for Adolphus D-50 (see Table 4.8) combined
with output from a 5/2/3 RASC run on 24 wells. As shown in Figure 9.12a
there are as many as 39 events on 28 levels in this well so that there is
good control in the vertical direction. The first derivative of the spline-
curve for SF = 2.2 (Fig. 9.12a) remains fairly constant. It has its largest
value at event level 14 (see Fig. 9.12b). This indicates that place where the
spline-curve in Figure 9.12a has its steepest dip. The pattern of Figure
9 . 1 2 ~suggests that rate of sedimentation was above average between
events 6 and 7 and also between events 21 and 22. These two maxima also
336

occur in Figure 9.12e which is the first derivative of the event-depth curve
(Fig. 9.12d), obtained by combining the spline-curves of Figures 9.12a and
9 . 1 2 ~with one another using the indirect method. The smaller peak i n
Figure 9.12e, which occurs at a depth of about 1600m, represents the place
(level 14) where the curve of Figure 9.12a has its steepest dip. The same

C EVENT LEVEL d AGE IN Ma

a 26 20 16 10 I 1 7 0 1 1 0 W 4 0 W S O l O O

e SEDIMENTATION RATE
f AGE IN Ma
10 D 0 7 0 6 4 5 2 1 0 70 W W 40 W 20 10 0

Fig. 9.12 Example of CASC displays for Adolphus well f5/2/3 RASC results using 24 wells). fa) Age-level
plot for SF=2.2. (b) First derivative corresponding to (a); note small peak near level 14. (c) Event level-
depth plot; note relatively steep slopes at depths near 0.7 k m and 2.2 km, respectively (d) Spline-curve
with small (default) SF fitted to combination of (a) and (c). ( e ) Sedimentation r a t e in 0.1 kmlmy
corresponding to fd); two relatively high peaks correspond to steeper slopes in (c); intermediate small
peak corresponds to highest first derivative in (b). (0 Event-depth spline-curve fitted directly to the data
using irregularly spaced depths; note similarity with spline-curve of (d); direct method yields poorer
results t h a n indirect method when one or more intervals between successive ages a r e much larger t h a n
average, due to high sedimentation rate or relative lack of microfossils.
 337

TABLE9.5

Information for Adolphus D-50 well used for CASC experiments of Figs. 9.12 to 9.14; ID are identification
numbers of foraminifers (cf. Tables 4.7 and 4.8); rank gives position of event in scaled optimum sequence;
age was derived from RASC distance; level refers to successive samples taken at different depths.

ID Rank Age Level Depth ID Rank Age Level Depth

(m.y.) (W (m.y.) (h)
10 3.0 6.634 1 0.318 147 23.0 35.561 15 1.622
71 8.0 17.695 2 0.400 - 260 27.0 37.942 15 1.622
16 5.0 12.354 3 0.435 60 34.0 43.353 16 1.662
18 7.0 15.312 4 0.482 32 33.0 42.457 17 1.731
20 11.0 18.557 5 0.574 40 38.0 42.477 18 1.767
201 12.0 20.393 6 0.854 30 42.0 44.916 19 1.804
26 9.0 19.465 7 0.903 49 43.0 47.339 20 1.860
15 13.0 21.542 8 1.086 - 29 31.0 41.523 20 1.860
- 81 18.0 28.791 8 1.086 90 46.0 48.912 21 2.996
- 69 16.0 27.683 8 1.086 - 37 49.0 49.778 21 2.996
24 17.0 25.444 9 1.250 93 48.0 51.331 22 2.285
- 33 24.0 35.964 9 1.250 36 47.0 49.189 23 2.383
- 202 19.0 29.272 9 1.250 164 55.0 54.274 24 2.414
259 21.0 30.452 10 1.323 so 56.0 55.003 25 2.487
- 25 20.0 28.015 10 1.323 - 230 51.0 55.046 25 2.487
263 28.0 40.263 11 1.360 54 58.0 56.640 26 2.525
82 22.0 34.115 12 1.470 57 45.0 49.428 27 2.567
85 29.0 39.358 13 1.479 - 56 59.0 59.009 27 2.567
- 261 26.0 38.057 13 1.479 55 60.0 61.158 28 2.622
203 37.0 40.481 14 1.616

three intervals with relatively steep slopes in the event-depth spline-curve

can be observed in Figure 9.12f, which resulted from applying the direct
method t o the age-depth values. Without further statistical analysis or
corroboration from other wells drilled in the immediate vicinity it is not
possible t o decide with certainty whether or not small fluctuations in the
rate of sedimentation, as shown in Figure 9.12e, are significant. Increased
smoothing in the event-depth diagram (Fig. 9.12d) would change the
pattern of Figure 9.12e much more drastically than the pattern of Figure
9.12d itself.

As was illustrated in more detail for Indian Harbour M-52 in Figures

9.9 to 9.11, minor smoothing in Figure 9.12a for Adolphus D-50 would, in a
multi-well comparison, only slightly change the position of isochrons in
Adolphus D-50. However, the widths of the error bars are proportional to
rate of sedimentation in both local and global error bar estimation and
these widths would change drastically if smoothing is increased. This is
because rate of sediment accumulation does depend strongly on choice of
smoothing factors.
338

9.4 Statistical selection of optimum spline-curves

In the preceding sections, extensive use of smoothing splines has been

discussed. In this respect, quantitative stratigraphy follows a general
trend in computer-based statistics where smoothing splines have become
widely employed a f t e r t h e i r invention (Reinsch, 1 9 6 7 , 1971;
Schoenberg, 1964) in the late 1960’s. The book by de Boor (1978) provides
an introduction to spline-fitting with computer programs in FORTRAN.
For comprehensive reviews of splines in statistics, see Eubank (1988) and
Wegman and Wright (1983). The approach t o spline-fitting taken in
mainframe CASC is that the user decides in a subjective way on a best
smoothing factor. It can be assumed that the latter lies somewhere
between the “default” value which is based on the law of superposition of
strata (age always increases in the stratigraphically downward direction)
and the straight line which represents the smoothest possible spline. An
additional guideline is provided by the value of the standard deviation
( = 0.7071) originally selected for events along the linear scale in
the RASC model. If all events in a well would occur at different levels, this
value (0.7071) would represent a good choice for the smoothing factor in
diagrams with RASC distance plotted along the horizontal axis. This
guideline applies only if all events, approximately, have equal standard
deviations as assumed in the scaling model.

The basic idea of the smoothing spline was explained in Section 3.1 1.
It was pointed out that S representing the sum of standardized residuals in
Equation (3.23) is distributed as chi-squared. This result is derived from
statistical theory for t h e distribution of t h e v a r i a n c e s2 (see
e.g. Hald, 1957, p. 278) which has mean E(s2) = u2 and variance
Var(s2) = 2u4/fwhere f = n-1. Setting S = ns2/u2, it follows that E(S) = n
and Var(S) = 2f. Thus the preceding interval extends from one standard
deviation below the mean ( = n )to one standard deviation above it. This
idea has led users of smoothing splines t o the choice S = n (“Reinsch’s
suggestion”, see e.g. Wahba, 1975). Because the smoothing factor is
defined as SF = (S/n)*,the use of Reinsch’s suggestion is equivalent to
setting SF = 1.0 if all values of s(yi) are known. This is in fact the method
of spline-fitting previously used for constructing geological time scales (see
Section 3.11) and in modified RASC (Chapter 8).
 339

9.5 Cross-validation method

Wold (1974) conducted a number of computer simulation experiments
for finding the optimum smoothing factor. In these experiments for which
all data had known standard deviations, setting SF = 0.84 to SF = 0.97
provided better results than SF = 1.00. This illustrates that, even if good
estimates of s(yi) are available, it is not known exactly which S F i s
optimum. For this reason, Wahba (1975) introduced the method of cross-
validation for experimentally finding the best smoothing factor. This
method has the additional advantage that it can be used even when
estimates of s(yJ are not available as in most applications of CASC.

Suppose that yi and si represent observed and fitted values,

respectively, and that residuals are written as Ri = yi-si ( z = 1, 2, ..., n).
Then:

(9.3)

In cross-validation, separate spline-curves are evaluated for m

different smoothing factors. Let s l ~ k( i = 1, 2, ..., n; j = 2 , 3 , ..., n-1; and
h = 1 , 2 , ..., m ) represent the ith value on a spline-curve for the hth
smoothing factor fitted to a reduced data set of size (n-1) obtained after
deleting the valueyj. Then, a cross-validation value CVk can be defined
as:

(9.4)

It is noted that this sum is based on (n-2)instead of n comparisons because

the first and last values, silk and snnk ( h = 1 , 2 , ..., m ) , are not available.
The optimum smoothing factor has the lowest value of CVk. In general,
many spline-curves must be fitted before a satisfactory solution to the
problem of optimizing SF is obtained. However, in our type of application
the number of spline-curves t o be fitted is not too large. For example, if
m = 30 and n = 22, the total number of separately fitted spline-curves
required for optimizing SF amounts to 600. Biostratigraphic datasets for
single wells often have n < 30 and m can be kept small by establishing a
range in which the optimum smoothing factor should fall before cross-
validation is applied. This range extends from its minimum value
340

corresponding to the law of superposition of strata (no decrease in age in

the stratigraphically downward direction) t o its maximum value
corresponding t o the SFvalue of the best-fitting straight line. The
minimum and maximum values themselves are possible solutions for the
optimum smoothing factor. For the preceding t w o reasons, cross-
validation generally yields good results requiring relatively little
computing time in biostratigraphic applications. It is not necessary t o
reduce the amount of computing time further by adopting one of the
approximation methods known as “generalized cross-validation’’ (Craven
and Wahba, 1979; Golub, Heath and Wahba, 1979; Utreras; 1981;
Silverman, 1984).

The method of cross-validation is part of CASC 2 in micro-RASC (see

Chapter 10). Table 9.5 showed CASC input for the Adolphus D-50 well.
There are 39 ages occurring on 28 separate levels. In the CASC run
previously applied to this dataset (see Fig. 9.12a), a spline-curve with
SF = 2.2 was fitted to 28 values, after averaging ages on the same level.
Table 9.6 shows results obtained by cross-validation applied to this
example. The first step for cross-validation in CASC 2 consists of
calculating the range for the optimum smoothing factor. The default
spline had SF,i, = 1.82 and the best-fitting straight line gave
SF,,, = 3.25, thus providing a range within which the optimum
smoothing factor should fall. In CASC 2 this range is divided into 10 equal

TABLE 9.6

CASC 2 output for Adolphus D-50 (age-level plot, cf. Fig. 9.13a); smoothing factors SFk range from
1.8158 for k = l (first spline-curve satisfying law of superposition of strata) to 3 2519 for k = l l (best
fitting straight line); optimum smoothing factor has lowest cross-validation value cvk.

1 1.8158 13.2344
2 1.9594 12.1967
3 2.1030 11.3395
4 2.2466 10.4125
5 2.3902 9.4796
6 2.5338 9.0214
7 2.6775 9.2514
8 2.8211 9.7706
9 2.9647 10.3320
10 3.1083 10.9710
11 3.2519 11.3223
 34 1

a AGE IN Ma b FIRST DERIVATIVE

ro M w UI 30 20 10

- -26

A SO 5.0883

C EVENT LEVEL d AGE IN Ma

5 0 2 6 ' 2 0 1 6 1 0 1 1 ~ o ~ m ~ ~ a o ~ i
10 ro

e SEDIMENTATION RATE
f AGE IN Ma
10 o I r e 6 1 s z 1 o ro m w u) so '20 10 o
to to

SF - 2.5800

I
SO - 2.5800

Fig. 9.13 Analysis on example of Fig. 9.12 for Adolphus D-50 repeated using optimum smoothing factors
obtained by cross-validation for spline-curves in Figs. 9.13a and 9.13f. Largest differences in fitted
curves occur in Figs. 9.13b (cf. 9.12b) and 9.13f(cf.9.120. For further explanation see text.

intervals and initially a cross-validation value is computed for each of the

11 equally spaced smoothing factors belonging to the range.
342

In Table 9.6, the minimum CV value occurs at SF = 2.53. A slightly

improved estimate of SF = 2.546 was obtaining after zooming in on the
vicinity of the initial minimum with a narrower range. Figure 9.13a
shows the spline-curve for this optimum smoothing factor. It is only
slightly smoother than the “subjective” spline-curve for SF = 2.2 shown in
Figure 9.12e. However, the first derivative of the curve for SF = 2.546
(see Fig. 9.13b) is considerably smoother than its counterpart for SF = 2.2
in Figure 9.12b. The optimum age-level spline was combined with the
same level-depth spline as before to give a new event-depth curve (Fig.
9.13d). This new spline closely resembles the previous result (see Fig.
9.12d). The new sedimentation rate (Fig. 9.13e) differs only slightly from
the old one in that the central and smallest of the three maxima in Figure
9.12e has disappeared.

Finally, Figure 9.13f shows the spline-curve for S F = 2.56

representing the optimum smoothing factor obtained by cross-validation
using irregularly spaced depth data. This new curve (Fig. 9.130 based on
the “direct” method is considerably smoother than its counterpart for
SF = 2.1 in Figure 9.12f. It is also much smoother than the new event-
depth curve (Fig. 9.13d) obtained by optimizing SF for the indirect method.
Although the optimum smoothing factors for the indirect and direct
method are nearly equal to one another, their corresponding splines turn
out t o be very different. It will be shown in the next section that the spline
of Figure 9.13d which is based on the indirect method is better than the
one shown in Figure 9.13e which is based on the direct method.

9.6 Jackknife method

Quenouille (1956) introduced the idea of splitting a sample of size n

(for independent and identically distributed random variables) into
g groups of size h ( n = g h ) , analyzing the data in such a way that (I) bias
would be redyed, and (2) a variance estimate would become available, for
&
an estimator 8 of a parameter 8 based on the sample of size n. Let ( i = 1,
2, ...,g ) represent the same estimator based on the i th reduced data set of
size (n-h).Then

(9.5)
 343

can be defined, leading to the jackknife estimator

,=l (9.6)

Tukey (1958) proposed that the g so-called pseudovalues ei

could be
treated as approximately independent and identically distributed random
variables in many situations. The statistic

(9.7)

then should have on approximate t-distribution with (g-1) degrees of

freedom. This would constitute a key statistic for robust confidence
interval estimation (cf. Miller, 1974). Considerable research has gone into
verifying and, by means of some counterexamples disverifying, the
usefulness of the jackknife for robust variance estimation. For a review,
see Efron (1982). The approach provides best results for ungrouped data,
i.e. if g = n ( h = 1). Wold (1974) reported that good results were obtained
by him applying the jackknife method for estimating confidence intervals
of parameters of spline-functions fitted to data.

Keeping the notation previously used for cross-validation, one can

define n pseudovalues for a spline-curve as

q, = ( n - 2 ) s L- ( n - 3 ) s
(9.8)

The subscript h has been dropped because a single value of SF is used in

each jackknife experiment. The pseudovalues lead t o the jackknife
values q iand their standard deviations B(q,):

n- 1

(9.9)

Consequently:

(9.10)
344

Setting t(n-3) = 2, this leads to the approximate 95 percent confidence

interval

Jackknife values can be obtained for all four coefficients which determine
a cubic curve for each the (n-1) intervals between successive values x L and
x L + l (i = 1, 2, ..., n-1). Use of all coefficients results in a jackknife spline
which interpolates between these successive values.

For example, Table 9.7 shows the values s L ,q Land s(qJ for the spline-
curve of Figure 9.13a. The corresponding jackknife spline based on
complete sets of four coefficients is shown in Figure 9.14a together with
95 percent confidence intervals for the values 9,. Comparison of the values
s(qi)in Table 9.7 indicates that spline and jackknife spline for SF = 2.456
are close t o one another. Nearly all standard deviations s ( q i ) are less
than SF. Only from level 10 to 14, the d q , ) values are relatively large as
can also be seen in Figure 9.14a. It would be possible to transfer the error
bars of Figure 9.14a to the data points in Figure 9.13d, and t o project them
along the depth scale by one of the methods illustrated in Figure 9.7.
Instead of expressing the uncertainty of the observed events with respect
t o their most likely positions, these new error bars would give the
uncertainty of the estimated ages themselves.

The jackknife method provides a valuable new tool for investigating

the validity of splines fitted by different methods. This will be
demonstrated by a comparison of the four jackknife splines shown in
Figure 9.14. Figure 9.14b is the jackknife spline for S F = 2.56
corresponding t o the spline of Figure 9.13f. This jackknife is for the spline
with optimum smoothing factor in the situation of irregularly spaced
depth data. It is fairly close t o its counterpart with subjectively selected
SF = 2.1 (Fig. 9.120 except near the top and bottom of the section where
the spline of Figure 9.13b would imply higher sedimentation rates. In fact,
the age is decreasing with depth in the lowest part of the section thus
violating the law of superposition of strata. This suggests that the method
of fitting a cubic spline with irregularly spaced data may not give
satisfactory results (also see experiments described in Section 3.6). Figure
9 . 1 4 ~with SF = 2.2 corresponds t o the spline of Figure 9.12a; and Figure
9.14d with SF = 2.1 to that of Figure 9.12f. The jackknife spline of Figure
9 . 1 4 ~does not closely resemble the spline of Figure 9.12a. The jackknife
 345

TABLE 9.7

CASC 2 output for Adolphus D-50 (age-level plot, ef. Figs. 9.13a and 9.144; the values s l are situated on
the spline-curve with optimum smoothing factor (SF= 2.456); the values q r with standard deviations
6 ( q , ) belong to the correspondingjackknife spline.

1 9.3954 6.3569 1.8090

2 11.6993 9.0325 1.SO35
3 13.9446 12.0333 1.3183
4 16.1968 15.2569 1.2735
5 18.5145 18.5826 1.2765
6 20.9281 21.9121 1.2336
7 23.4611 25.0837 1.1596
8 26.1030 27.9331 1.3708
9 28.7515 30.5142 2.0336
10 31.2879 32.9016 2.7904
11 33.6144 35.1614 3.3494
12 35.6327 37.4329 3.4880
13 37.3720 39.5062 3.2675
14 38.8722 41.2575 2.7789
15 40.2034 42.6240 2.1674
16 41.4596 43.5825 1.6073
17 42.6775 44.3336 1.2058
18 43.9157 45.0106 1.0935
19 45.2303 45.6888 1.1591
20 46.6426 46.4389 1.2063
21 48.1466 47.3135 1.1650
22 49.6923 48.4075 1.1132
23 51.2528 49.7226 1.0849
24 52.8331 51.1675 1.0637
25 54.4094 52.7214 1.1012
26 55.9815 54.3178 1.2085
27 57.5750 55.8434 1.3380
28 59.2145 57.2 140 1.4501
-

spline for SF = 2.2 is not as smooth as the spline for SF = 2.2 that was
originally selected in a subjective manner.

Although there is little difference between the splines of Figures

9.12a and 9.13a, the difference between the corresponding jackknife
splines of these patterns suggests that only the spline for the optimum
smoothing factor (SF = 2.456) is satisfactory. The jackknife spline of
Figure 9 . 1 4 ~also violates the law of superposition in several places. Even
more severe violations of this law can be observed in Figure 9.14d where
the jackknife spline dips in the wrong direction in several places. These
346

2 F - I8 19

~ 68
58 i- 1Age i n L

81" ,
5
18
15 1
d

I
68 58 48
Age in k
38 28 18 8

j.j
25

t38 /3

Fig. 9.14 Jackknife spline-curves with approximate 95% confidence limits for Adolphus D-50 results
previously shown in : a. Fig. 9.13a; b. Fig. 9.13f; c. Fig. 9.12a; d. Fig. 9.12f. The optimum smoothing
factor patterns of Figs. 9.13a and 9.13f a r e relatively closely approximated by their jackknife splines,
contrary to the subjectively selected spline-curves of Figs. 9 12a and 9.12f. The latter two jackknife
splines (Figs. 9 . 1 4 ~and 9.14d) show violations of the law of superposition of strata. In general, the
indirect method (Figs. 9.14a and 9 . 1 4 ~ yields
) results which are superior to those of the direct method
(Figs. 9.14band9.14d).

are characterized by lack of data along the depth axis due to relatively
high sedimentation rates. Although the subjectively derived event-depth
curve for SF = 2.1 (Fig. 9.12e) is relatively close t o the "optimum" event-
depth curve shown in Figure 9.13d, it obviously could not be duplicated by
its jackknife estimator. This confirms that it may be dangerous t o fit
splines to irregularly spaced data. The results of Figure 9.14 clearly
demonstrate that the indirect method of constructing event-depth curves
illustrated in Figure 9.6 is to be preferred to the direct method. The
discrepancy between the patterns of Figures 9.13d and 9.13f also can be
explained now. Although the pattern of Figure 9.13f is for a n optimum
smoothing factor and was reasonably well duplicated by its jackknife
spline (Fig 9.14b), the irregular spacing of control points along the depth
 347

axis resulted in a pattern which is too smooth in comparison with the

pattern (Fig. 9.13d) obtained by the indirect method.

It is noted that the experiments with the CASC 2 computer program

described in this section made use of the method of assigning equal
weights to ages for all levels. It is possible in CASC 2 to assign more
weight to level values based on more than a single age. This alternative
procedure was applied but it led t o results which are not markedly
different from those described in this section.

9.7 Computer simulation experiment for event-depth spline fitting

with error analysis
During development of the RASCKASC procedures, three criteria of
method evaluation were employed: (1) the method should have a firm
stratigraphic foundation; (2) it should be logically coherent from a
mathematical-statistical point of view; and (3) the computer programs
should be efficient as well as user-friendly.

The first aim (1) is promoted by systematic comparison of RASC

zonations independently obtained and correlations with subjective results,
and by evaluation of computer outputs by stratigraphers. Computer
simulation experiments are helpful in (2) and (3). Results for one such
experiment are displayed in Figure 9.15 and Tables 9.8 and 9.9. A
+ +
theoretical age-depth curve t = 9.155 0 . 6 8 5 ~ 0 . 1 6 5 ~- 0~ . 0 0 0 5 ~(unit
~
of x = 100m) is shown in Figure 9.15a. Twenty-one random normal
numbers Eoi (zero mean, unit variance) of regularly spaced points
(labelled i = 1, 2, ..., 21 in Table 9.9) were added in order to simulate
observations, yi, of biostratigraphic events in a hypothetical Cenozoic
basin. In Reinsch - De Boor spline-fitting, the smoothing factor (SF)fully
determines the shape of the best-fitting spline-curve. In CASC 2 (see
Chapter lo), SF is input and, as elsewhere in this book, has the following
meaning: the square of SF is equal t o the averaged squared deviation
2
(yi-si) between observations yi and their corresponding values si on the
spline-curve s.

An optimum value of SF can be found by cross-validation (see Section

9.5) as illustrated in Table 9.8. In general, SFmi, 5 SFOpt5 SF,,,
where the minimum smoothing factor (SFmi,) represents the first-fitted
smoothing spline curve whose age increases monotonically with depth.
348

Age in Ma Age in Ma
0 40 30 20 10 0 40 30 20 10
0 0
4

m
2 0
E E
X X
._
C c
5a 5
a

'4
0
W
d
r r

K P
N

Age in Ma Sediment Accumulation Rate

0 40 30 20 10 4 3 2 1 0
0 0
D
i

m
2 0
E E
Y Y
.-C c
5a 5
a
?L 0
'0 W
r r

: P
N

Fig. 9.15 Computer simulation experiment. Random normal deviates were added to theoretical curve
(A). Cross-validated smoothing spline (B) was approximated by i t s jackknife estimate (C). First
derivative of spline-curve ( B ) gave sediment accumulation rate curve (D, solid line) which is compared
with first derivative of theoretical curve (D, broken line).

The maximum smoothing factor (SF,,,) corresponds to the best-fitting

straight line. The optimum smoothing factor (SF,,t) has minimum cross-
validation value within the interval between SFmi, and SF,,,. The
spline-fitting routine in CASC 2 is a modified version of De Boor's (1978)
FORTRAN program which uses the secant method, returning a smoothing
factor that is slightly greater than an input SF. For this reason, 1.170 was
selected from Table 9.8 as input for the cross-validation smoothing spline
with SF = 1.174 shown in Figure 9.15b. Obviously s in Figure 9.15b is
close t o t i n Figure 9.15a. In practice, t is not known and one would like t o
 349

TABLE9.8

C A S C 2 output for computer simulation experiment of Fig. 9.15. See Table 9.6 for explanation of column
headings.

k SFk cvk
1 0.828 3.607
2 0.897 3.314
3 0.967 2.968
4 1.036 2.565
5 1.106 2.219
6 1.175 2.131
7 1.245 2.192
8 1.314 2.290
9 1.384 2.395
10 1.453 2.445
11 1.523 2.448

construct confidence intervals on s t o evaluate the difference between s

and (unknown) t, and to check residuals (yi-si) for possible outliers. This
problem has been studied by Wahba (1983) and Eubank (1984) who have
shown that the column vector S of values si are related to column vector Y
for yi by S = HY where the h a t matrix H has properties similar to the h a t
matrix in regression analysis. Several methods have been developed for
obtaining H in explicit form. This allows estimation of the following two
variance-covariance matrices: Var (si-ti) = u 2H , and Var (yi-si) = u 2 (I-H)
where I is the identity matrix.

In this section, the following procedure is used for obtaining the

diagonal elements hii of H . S was closely approximated by its jackknife
spline-curve q with values qi a t the observation points (see Fig. 9 . 1 5 ~and
Table 9.9). The jackknife method provides variances s2(qi)of the values qi
which are approximately equal to u 2 h2i l being the diagonal elements Var
(si-ti). The smoothing factor SF = 1.174 estimates u. The diagonal
elements of Var (yi-si) can also be estimated because qi = si. These are
written as s2(Eli)in Table 9.9. Approximately 95% and 99% confidence
intervals are obtained by multiplying s(E1i) by 2 and 3, respectively. As
expected, no outliers are indicated for the random normal numbers used in
t h e computer simulation experiment. Because s ( E 1 i ) applies to
350

observations used for estimating the spline-curve s, Table 9.9 also shows
residuals E2i = y2i - qi for new observations y2i obtained by adding 21
other random numbers to t i . These new observations have wider
confidence belts with widths controlled by s(E2i) = SF d ( l + h i i 2 ) , also
shown in Table 9.9. This second type of confidence belt would, for example,
apply to test suspected outliers not used for calculating the smoothing
spline.

The three standard deviations s(qi),s(E1i)and s(Egi) can be projected

onto the depth axis by using the sediment accumulation rate curve (Fig.
9.15d). It is noted that this type of curve is more difficult to estimate than
the event-depth curve as indicated by larger discrepancies in Figure 9.15d.

TABLE9.9

Random normal deviates (Eo,) were added to theoretical values (1,) on curve of Fig 9 15a to give observed
values y, Cross-validated smoothing spline values (s,)on curve of Fig 9 15b were approximated by
their jackknife estimates (qJ of which standard deviations dq,) could be computed. Standard deviations
s (El,) and s (E2J are for residuals of data used ( E l , ) and not used ( E z J for estimating qL,respectively

1
-
1 1000 054 10 54 9 69 9 79 102 075 059 048 155

2 I 1 15 063 11 77 11 14 11 23 062 054 1 00 0 57 133

3 12 56 1 99 10 57 12 66 12 64 0 57 208 102 091 131

4 14 22 1 66 15 87 14 29 I4 03 0 71 185 097 110 137

5 16 08 I 22 14 86 16 05 I5 59 084 073 082 0 46 144

6 18 13 0 04 18 08 17 97 I? 39 090 069 076 140 1 48

7 2032 056 19 76 20 03 19 46 090 030 075 112 1 48

8 2263 036 23 00 22 24 21 82 087 117 079 162 146

9 2504 245 22 59 24 58 244; 078 181 0 88 171 141

10 2750 0 16 27 67 27 04 27 04 0 59 0 62 102 0 15 131

11 3000 0 15 29 85 29 55 2973 0 45 0 12 108 073 126

12 3250 0 13 32 63 32 04 32 39 0 51 0 24 1 06 151 128

13 3496 2 33 32 63 34 47 34 90 070 227 094 042 I 37

14 37 37 130 38 66 36 81 37 17 0 91 149 075 056 I48

15 3968 075 40 43 38 96 39 32 0 94 121 0 70 I72 151

16 41 87 151 40 36 40 93 41 36 079 100 087 0 85 141

17 43 92 174 42 18 42 78 43 21 0 59 104 102 147 131

18 4579 110 44 68 44 54 44 87 0 45 0 19 I08 1 82 126

19 47 44 204 45 40 46 26 4642 0 44 1 02 109 076 125

20 4886 067 49 52 47 93 47 95 0 55 151 104 0 43 129

21 5000 152 48 48 49 55 4958 078 110 088 026 141

 35 1
9.8 Regional application of RASC and CASC
The geological use of the RASC/CASC method and its value in
sedimentary basin analysis will be illustrated by means of examples
drawn from exploration micropaleontology. The first examples are based
on the original distribution of 168Cenozoic Foraminifera in 2 1 Grand
Banks and Labrador wells (cf. Fig. 6.2), and of 116 taxa of Mesozoic
foraminifers in 16 Grand Banks wells. The latter databank was largely
prepared by Williamson (1987). Later examples will deal with the North
Sea Basin and integration of different families of microfossils
(foraminifers; dinoflagellates; spores and pollen). The Cenozoic databank
used for running CASC on the Labrador Shelf of Grand Banks is the same
as that previously used for RASC (Chapters 4-6), with the one difference
that in the Bjarni H-81 well, taxa 54 and 55 (Gavelinella beccariiforrnis
zone, Paleocene) have been added (cf. Table 4.8). Both taxa were observed
at a depth of 6660ft.

The discussion in the remainder of this chapter deals with the

following questions: (1) what is the stratigraphical meaning of
RASC/CASC-type of correlation?; and (2) what is the degree of confidence,
expressed in depth and in linear time units of CASC correlation, and to
what extent is the error bar useful for geological interpretation?

In order t o find answers t o these questions, conventional or subjective

and more objective CASC-type correlation of wells are compared using
three related stratigraphic criteria: ( a ) selected zone markers;
(b)assemblage zones; and (c) isochrons in Ma. In all examples, the
underlying biostratigraphic zonations were defined with the RASC
method, but in principle CASC can be applied to non-RASC zonations
based on biostratigraphic and other stratigraphic events.

Firstly, ten selected zone markers were traced through six wells.
Starting point was the Cenozoic scaled optimum sequence (Fig. 6.2, 7/2/4
run for 21 wells). For the interactive spline fitting of the bivariate plots,
all CASC defaults were accepted, unless otherwise specified. The first
CASC default is the smoothing factor (SF)that defines the spline-curve for
which an increase in position or depth along one axis does not anywhere
correpond t o a decrease in position (or time) along the other axis. This
default is obtained by means of an algorithm that calculates spline-curve
fits with SF increasing o r decreasing according to a binary search method.
352

The default satisfies the condition that the observed sedimentation rate is
never negative.
In three wells the mainframe CASC cursor option was used to delete
aberrantly positioned events: in Hibernia P-15, one point was deleted on
event level 12; in Bonavista C-99, three points were deleted (on levels 4 , l l
and 15), and SF for the events versus depth graph was changed from its
default (=0.02) to 0.15; and in Snorri J-90, one point was deleted on
level 6.
The results of the CASC multi-well comparison are shown i n
Table 9.10, listing both the observed and the most likely depths of the ten
selected Paleocene through Miocene zone markers 50 (561, 90, 32, 29,261
(260), 259, 24, 26(15), 18 and 16. In two wells substitute taxa were
correlated rather than the three designated events50, 261 and 26. The
substitutes 56, 260 and 15 are neighbors of the original events in the
optimum sequence.
In most instances, the observed and the most likely depth values are
within half the length of the error bar (68% probability) around the most

TABLE 9.10

Observed (above line) and most likely depth (in m) of ten Eocene through Miocene zone markers in six
wells. The fossil numbers a r e the RASC dictionary numbers. Results are based on optimum sequence
CASC (21 wells; k,=7, rnCl=2,rn,2,4); * means that a t that site substitute fossils (neighbors in the
optimum sequence) were used.

Bibernia Molpbw &mavista I. mrbour Bjarni brri

P-15 D-50 c-99 Il-52 8-81 J-90
CeratobuliriM cantraria - 16 310 485 887 750 872 1262
350 t 85 1130 t381 571 t132
s p i r o p 1 e e t u i a a carinate - I8 275 512 2030 649 1085 1701
334 t 65 481 t 78 1681 f363 673 t126 1025 t 18 1533 t167
b i g e r i n a dumblei - 26( 15) 550 933 2377 1261
619 f330 726 t228 2059 I162 786 t155* 1094 t 14 1756 t 42
~ r r i l i o aa l s a t i e a - 24 1075 1280 2377 2097 1298
1000 t 75 1164 +I15 2372 t13 1344 t532 1307 t 31 1910 t 29
-i.cua latus - 259 1125 1153 2316 1646
IU83 t 45 1291 t 71 2571 t135 1655 t147 1490 t 27 1983 t 26
qlophr.g.oides dteri - 261(260) 1185 1509 3078 1704 1634 213
1176 t 40 1471 f171 2889 f215 1864 t404* 1577 t 16 2067 t 22*
Cgeluina aplectens - 29 1195 1890 3109 2396 1634
1721 f 95 3099 t 17 2212 t489 1635 t 14
dD8phaeroidina sp. 1 - 32 1125 1761 3386 1935 1695 2112
1761 t 62 3144 t 99 2275 ,266 1653 t 15 2155 t 27

Acarioioa deoss - 90 2062 3478

2293 t249 3423 t 96 2528 t497 1763 t 8 2651 t 59
Subbotioa p t . g a o i e a - 50(56) 2517 2914 2009 2932
2501 t 66 2861 t327* 1812 t 15 2798 t 51*
 353

Fig. 9.16 Tracing of ten foraminifera1 events through six wells, using the CASC (optimum sequence)
method to calculate the most likely depths. Black bars show the deviations of these depths from the
observed depths. The chronostratigraphic segmentation is based on observed depths only.

likely value. As pointed out in the previous section, the actual precision of
the estimated depth of a n event in a well is probably greater t h a n that
indicated by the local error bar for single event positions along the spline
curve. Also, the local error bar at any depth is initially calculated over the
time interval along the (scaled) optimum sequence scale ( y ) , as defined by
twice the standard devition (SD) in t h a t ( y ) direction. I t is directly
354

proportional t o the fitted average sedimentation rate for each point (cf.
Fig. 9.7).

Figure 9.16 graphically correlates the ten events through the six
wells. The conventional chronostratigraphic segmentation, which is
shown for comparison, only uses the observed depth of events. The new,
most likely, zone marker depths would lead t o slight up or down
adjustments of the age boundaries. It could be assumed that such a change
might violate stratigraphic boundaries as adjusted for major lithology
changes as determined from well logs. However, using sonic and gamma
logs, no evidence for this was found. In the Snorri well there is no direct
micropaleontological evidence for the presence of events 259, 24 and 26,
associated with Oligocene-Early Miocene strata, although the CASC
method predicts their likely depth in this well. These depths are not
unreasonable given that Oligocene strata were thought t o be present at
that depth, based on palynology.

The conversion of the scaled optimum sequence t o a (local)

biochronology enables the stratigrapher familiar with CASC to trace
isochrons in the same way as zones were traced. The procedure begins
with the designation of numerical ages in Ma t o those events in the scaled
optimum sequence for which literature-based ages are available. In all,
23 events were dated this way, as explained below. The time scale is that
of Berggren et al. (1985). The regional use of the standard planktonic
zonation is as in Gradstein et al. (1985) who followed Gradsteinand
Srivastava (19801, and Gradstein and Agterberg (1982):

63 Ma - events 253 and 61 - S u b b o t i m triloculinoides and S . pseudobulloides - two rare events

that mark approximately the end of Danian time.

58 Ma -event 55 - Gavelinella beccariiforrnis; occurs up to standard zone P5 (Tjalsma a n d

Lohman, 1983), which fits well with i t s disappearance in t h e Adolphus well together with
Arugonia uelascoensis (Paleocene) and below the appearance of Pseudohusligerim (post P5).

57 Ma - event 194 -Planorotalites chaprnani; disappears in standard zone P6. Specimens a r e often
transitional between P. chaprnani and Pseudohastigerina. The latter is thought to appear at the
boundary of P5 and P6, or k 57 Ma ago.

55 Ma-event 50-Subbotina putugonica; is frequent in the Ypresian of Belgium (Muller a n d

Willerns, 1981), in the Moe Clay of Denmark and in the Lower Eocene of the North S e a a n d
Labrador Sea The end of the S. patagonica peak occurs a t the boundary of N P l l N " 2 , which
 coincides with the boundary of the Morozouella formosa formosa Zone, a t the time of Anomaly 24,
just after 55 Ma.

52 Ma-event 93 -Acarinina broederrnanni; the species has its top well below A . densa, probably
in the A . pentacamerata - Hantkenina aragonensis Zone, near the Early-Middle Eocene boundary
a t 52 Ma. In some RASC runs, A . broedermanni falls between Early and Middle Eocene zones.

49Ma-event 90-Acarinina densa; this is the time of the optimum climatic warming in the
Labrador Sea, in early Middle Eocene time. Less common a t this time a r e A . senni, A . aff.
penlacamerata, A . aff. broedermanni, Mororouella caucasica, M . spinulosa, a n d M . a f f .
aragonensis. The event probably falls in the Hantkenina aragonensis - Globigerinatheka
subconglobata Zone a t Anomaly 21 time or 52-46 Ma (average 49 Ma).

(7) 40 Ma - event 29 - Cyclamminu arnplectens; in RASC runs this event falls below Turborotalia
pomeroli and Globigerina yeguaensis and above Acarininu densa. In Poland its peak occurrence is
so-called Middle Eocene; it is less frequent in upper Eocene strata (Gradstein and Berggren, 1981).
Theevent was tentatively placed a t 40 Ma.

(8) 38 Ma -event 85 - Pseudodhastigerina micra; same reasoning a s for Turborotalia pomeroli (see
below), but often disappears in slighty older beds, a s also shown in the scaled optimum sequence
(Fig. 6.2).

(9) 37 Ma - event 33 - Turborotalia pomeroli; co-occurs in southern wells with Subbotina linuperta,
Globigerinu yeguaensis and Pseudohastigerinu micra, of the Turborotalia cerroazulensis Zone, late
Late Eocene. The top was placed just below the inferred Eocene/Oligocene boundary.

(10) 28 Ma - event 24 - Turrilina alsatica; the top of this distinctive Oligocene taxon roughly equates
with the top of the Boom Clay in Belgium and the top of the Globorotalia opirna opima Zone, a t f
28 Ma.

-20 Ma - event 26 - Uuigerina dumblei; slightly older than Globigerina praebulloides

20 Ma - event 137 - Globigerinoides primordius trilobus; rare Early Miocene event

17 Ma - event 15 - Globigerina praebulloides; disappears locally with Sphaeroidinella seminula

and with or just below Globorotalia scitula praescitula, which may equate with the G . fohsi
peripheroronda Zone, e a r l y Middle Miocene of S c o t i a n S h e l f wells ( G r a d s t e i n a n d
Agterberg, 1982). The RASC run 7/2/4 indicated an average disappearance in t h e Uuigerina
dumblei zone. Its local extinction was placed between 14 and 20 Ma (average 17 Ma).

15 Ma- event 179 -Globorotalia scifula praescitula; probably occurs in the late Early to early
Middle Miocene warming event, a s observed from the northern incursion of warmer water
planktonic taxa.
 3.5 Ma -events 266,4,269 and 5 - Both Globorotalia puncticulata, G. inflata, G. crassaformis, and
Neogloboquadrina atlantica are thought to disappear with the onset of major glaciation in the
Labrador Sea, dated at approximately 3.5 Ma.

Four other events occur a t or near significant breaks in the 5-2-3 and 7-2-4 scaling solutions for
21 and 24 wells. These breaks were equated with zonal boundaries and series breaks as follows:

58 M a - event 56 - Glomospira corona; Paleocene-Eocene boundary on (upper)continental margin

wells.

52 Ma - event 57 - Spiroplectammina spectabilis LCO; Early-Middle Eocene boundary;

LCO = Last Common Occurrence.

37 Ma - event 259 - Ammodiscus latus: Eocene - Oligocene boundary.

11 Ma - event 17 - Asterigerinagurichi; Middle-Late Miocene boundary.

Figure 9.9a was a plot of the ages of the previously listed events in a
RASC distance scale (21 wells, 7/2/4 run) versus linear time scale.
Smoothing of the spline-curve function diminishes some of the uncertainty
in subjective assignment. The spline function now can be used t o convert
the RASC distance scale into an age scale.
Next, the question can be asked what is the most likely depth in the
wells of the principal boundaries between RASC zones, expressed in Ma.
Gradstein and Agterberg (1985) have traced the boundaries between the
successive Cenozoic RASC zones (Fig. 6.2), which are close approximations
to the boundaries between Paleocene and Eocene (-56 Ma), Early Eocene
and early Middle Eocene (-52 Ma), early Middle Eocene and Middle Eocene
(-49 Ma), Middle and Late Eocene (-44 Ma), late Eocene and Oligocene
(-36 Ma) Oligocene and Miocene (-24 Ma), Early and Middle Miocene
(- 16 Ma) and top of Middle Miocene (- 12 Ma).

Table9.11 lists most likely and subjectively determined (as far as

known) depths for these isochrons. The same results are plotted in
Figure 9.17, with the wells arranged latitudinally (48"- 58"N). The CASC
depths are from a batch run that accepted all SF defaults. Although it
yields more crude results than interactive runs, this procedure takes less
time and the actual depth estimates are not influenced much. As
explained in the previous section, the choice of SF has much more
influence on the average rate of sedimentation and hence on the error
estimation, than on the actual depth of the isochrons. In a few instances,
default smoothing yielded unacceptably steep spline fits, and the local
 357

TABLE 9.11

Observed (above line) and most likely depth (in m) of the 5 6 , 5 2 , 4 9 , 4 4 ,36, 24, 16 and 12 Ma isochrons in
10 wells on the Grand Banks and Labrador Shelf. Results are based on scaled optimum sequence or
distance-CASC(21 wells; k c = 7 , rncl=2, mc2=4).

error bar estimate was deleted. In one well, Karlsefni H-13, both
foraminifers and palynomorphs agree on the absence of Oligocene beds
(Turrilina alsatica Zone). Batch CASC calculates a thin Oligocene
interval (24-36 Ma). Above the Eocene, the well has only a few data points
and results are crude.

The local error estimates of the most likely depths for the isochrons
are within 1 t o 10% of the actual depth values, and more frequently 2 to
5%. In about ten cases the subjectively assigned depths for the zonal
boundaries as converted to isochrons are outside the 68%confidence limits
(k1 SD). For geological interpretations, it should be borne in mind that
the error in most likely depth is an upper limit, and the SD is probably
smaller by a factor that, amongst others, is related to the number of
observations per spline-curve, as explained earlier. Palynologically
determined depths for these stage boundaries often are outside the depth
interval (most likely depth k 1 SD), calculated by CASC. The errors in
this independent biostratigraphic correlation are unknown, but the
comparison suggests that multiple biostratigraphy uncertainties exceed
the CASC-type of errors using one fossil discipline only. The conclusion
may be drawn that the CASC program is able to predict reliable and
objective well to well isochrons. The error expression, that remains vague
in conventional, subjective correlation schemes, is conservatively large
when one fossil discipline only is used.
358

I GRANOBANKS I LABRADOR SHELF I

Hibernia F Foam Adolphur Domlnlon Banavisla I Harbour Gudrld Blarnl Snorrl Karlreh
12 Ma
M ddle MloCenC

Fig. 9.17 Correlation of 8 Cenozoic isochrons, according to their most likely depths in 10 wells on the
Grand Banks and Labrador Shelf. The depths were computed by means of the RASC-CASC method
explained in the text. Subjective estimates fur the depths of these isochrons a r e shown with x.

9.9 Application of RASC and CASC to the Hibernia Oilfield

Williamson (1987) has investigated the application of ranking,
scaling and correlation in time t o the Mesozoic microfossil record
recovered from the Grand Banks, particularly for the area centered around
the Hibernia oilfield. Figure 1.1 illustrated the cumulative frequency
distribution of the highest occurrences of 116 taxa of foraminifers in
13 deep wells based on Williamson's original data (Fig. 9.18). This dataset
later was enlarged to encompass the Upper Jurassic and Lower Cretaceous
microfossil record of up t o 25 wells. Comparison between Williamson's
original zonation using thresholds of h, = 4 and rn, = rn,l = n , 2 = 3
which leaves 54 events, and subsequent runs of h, = 7 , m,l= 1 and
n c 2 = 4 (or 5) using 18 t o 25 wells, gave close to 50 events with virtually
 359

t I WEST
FLYING FOAM
-I L-23

HlBERNlA
OFLYING FOAM

NAUTILUS C-92

K-1800B-08
0
ADOLPHUS D-50
+

G-55 000350P-15

OBEN NEVlS 1-45

HEBRON.l-13

OEGRET K-36

-
0 Km 20

Fig. 9.18 Locations of 13 boreholes used by Williamson (1987)for RASCXASC application on northern
Grand Banks.

the same zonation. For this reason, the concise account of Williamson
(1987) is followed with minor changes, using his original illustrations.
F i g u r e 9.19 i s t h e s c a l e d o p t i m u m s e q u e n c e w i t h
chronostratigraphically useful average interval zones highlighted through
shading. Based on the original RASC run with 54events, t o which
9 unique events were added for (further) chronostratigraphic calibration,
eleven RASC zones s t a n d out, numbered from X I t h r o u g h I ,
Kimmeridgian-Cenomanian. This zonation considerably expands
stratigraphic resolution previously a v a i l a b l e . T h e r e i s good
correspondence of the average position of the disappearance levels of the
taxa in the wells and the upper part of stratigraphic ranges reported in the
literature. Some longer ranging taxa of the literature, on the Grand
Banks have relatively short ranges, as is the case with L. nodosa (no. 10)
and D.gradata (no. 111). N . uarsouiensis (no. 64) was not previously
reported so young. The tight clustering of events in the Albian zones 111
and I1 reflects considerable uncertainty on their exact disappearance
360

Fig. 9.19 Williamson’s (1987) eleven-fold average interval zonation, using ranking and scaling for the
Upper Jurassic and Lower Cretaceous foraminifera1 record, northern Grand Banks. Asterisks indicate
unique events.
 361

levels. For example, P. burtorfi is considered t o disappear later than R .

ticinensis, but in the zonation the order is reversed. It turns out that
R. ticinensis is rare which leads to poor sampling for the event and that the
more common P. buxtorfi in the wells is associated with other taxa of
“older but less certain” stratigraphic position. Zones I1 and 111, therefore,
reveal strong overlap in age.
As reported earlier, large distances between successive interval zones
in the scaled optimum sequence are caused by major sedimentary changes
or breaks that separate the majority of events below from those occurring
above it. Figure 9.19 shows large interfossil distances between zones X
(Tithonian) and IX (Valanginian) and between zones VI (Barremian) and
IV (Aptian). The lower of the two breaks is mainly due to the nonmarine
or very shallow marine facies probably of Berriasian age, which has a
paucity of shelly microfossils. This break also may be associated with a
condensed limestone sequence (seismic marker A), and may be related to
changes in sea level also observed in Portugal. The younger of the two
large breaks is the so-called pre-Aptian unconformity associated with
RASC zone V, below seismic marker B.
Because events in the RASC zonation are present in at least 4 wells,
well to well correlation is relatively easy. Williamson (1987) executed
both a subjective (manual) and an automated correlation. In the former
exercise, boundaries of zones were placed with reference t o the order of
events in each well, and any event t h a t clearly did not fit in, or
accompanied a group of events it was not associated with in the scaled
optimum sequence of Figure 9.19, was given less weight or ignored.
The quantitative correlation framework is based on the scaled
optimum sequence. The upper part of Figure 9.20 shows depth values of
RASC zones in each well. The numbers above each boundary are from a
subjective interpretation of RASC results. Numbers underlying each
boundary are the most likely depth for each zone calculated, using CASC.
Numbers in parentheses are the local error bar estimates in meters below
and above the CASC depth. As may be seen, the subjective zone depths
generally are within the error ranges calculated. Although there is no
easy choice between right or wrong in geological correlation, the close
match of the two types of correlation is a means of model verification.
The next step is t o convert the scaled optimum sequence of
Figure9.19 t o a local time scale, using the ages in millions of years of
362 0
0
I
I
E
N

I
DEPTH(m1
0
0

I
0
P

Fig. 9.20 Upper part: Depth values of RASC zones in northern Grand Banks wells. Numbers above
each boundary a r e based on subjective interpretation. Below each boundary a r e most likely depths using
the CASC method with error bars in meters in parentheses. Lower part: Comparison of subjective
(solid line) and most likely (dashed) depths for Cretaceous isochrons in northern Grand Banks wells
(after Williamson, 1987).
 363

several good marker events. Each (CASC) age versus depth plot per well
was executed with isochron boundaries for zones and the result is
displayed in the lower half of Figure 9.20. The dashed lines are based on
the CASC method and the solid lines are a subjective interpretation. An
advantage of the CASC type of interpolation method is that it can be used
for isochron cross-sections at for example 1m.y. intervals. Such cross-
sections as constructed by Williamson (cf. Williamson and Agterberg,
1990) have realistic geological properties and are of use in relating seismic
cross-sections to geochronologic results and in detection of hiatuses in one
or more wells. This type of application considerably enhances the role of
biochronology in regional basin studies.

S o far, the examples of automated correlation involved Cenozoic

foraminiferal events, zones and isochrons based on the RASC zonation in
Labrador and Grand Banks wells, and Early Cretaceous isochrons based
on the RASC zonation in the Hibernia oilfield, off Newfoundland
(Williamson, 1987). Previous analysis based on subjective age-depth data
consistently confirmed results obtained by the CASC model. The error for
the most likely depths of the correlation lines rarely exceeds 10%; it
commonly is 2 to 5%. CASC-type of age/depth data offer the potential for
significant contributions t o analytical error analysis i n tectonic
subsidence and sedimentation calculations. RASC and CASC make
subsidence analysis more objective and accurate, and easier t o perform by
non-paleontologists.

The procedure used t o derive the objective schemes depicted in

Figures 9.19 and 9.20 involved several steps. As has been seen, a
prerequisite for the derivation of the correlation scheme is the successful
application of the quantitative RASC program. This information was
made use of by CASC together with additional information in DEP files
that introduced recorded depth values (in meters) of each event in each
well and age estimates of selected taxa (in Ma) from the scaled optimum
sequence. The end result was a sequence of objectively derived isochrons
plus standard errors.

As pointed out in Williamson and Agterberg (1990),the application of

CASC to the foraminiferal data set of the Hibernia area enables a more
precise chronological framework within which t o consider the
relationships of particular sandstone bodies, especially those bodies of
economic interest. Early Cretaceous and Late Jurassic sedimentation in
the study area resulted in the accumulation of a thick sandstone-shale
364

sequence in a fluvio-deltaic setting which includes the Hibernia “Giant”

oil field. Precise determination of the temporal interrelationships of the
economically important sandstone sequences in the Hibernia area are
depicted on Figure 9.20 (upper part). The Avalon sandstone member
represents the youngest reservoir unit in this area and is thought t o
represent shoreline sand deposits (McKenzie, 1981). This sandstone lies
within RASC zone IV and is closely associated with the 115 Ma isochron
(mid-Late Aptian). CASC isochrons 105-115 Ma are “missing” o r
extremely condensed in some wells; for instance, in Hibernia B-08. Figure
9.20 shows how the chronologic position of this sand body fluctuates
indicating a degree of diachroneity. The main Hibernia sand is markedly
associated with RASC zones IX and X (Fig. 9.19) and isochrons 141-148
Ma (Fig. 9.20); i.e. from the data examined in this study the Hibernia sand
sequence seems to straddle the Jurassic-Cretaceous boundary.

The results and discussions of applications of RASC and CASC in this

and previous sections have a demonstrable reproducibility and
furthermore allow experimentation of results using different threshold
levels. Thus detailed interpretive scrutiny of results and the steps
required t o obtain them is possible. In addition, the methods allow
development of final interpretations that allow easier communication in a
scientific way t o fellow workers. Biostratigraphers then are able to
express numerically the uncertainty accompanying their zonation and
correlation schemes. Other benefits such as the ability to deal with ever
expanding databases, graphic display and data input and retrieval also are
of significance. Of greater implication, however, is the potential
contribution t o basin history analysis. The following two examples serve
to illustrate this point in more detail.
Burial history or subsidence curves can be derived and backstripped
by computer t o investigate the relative effects of sedimentary loading,
eustacy, paleobathmetry and tectonics upon the geohistory of a n area.
Previously, the regional time scale or rather the biozonation which
provides the chronostratigraphic control had been derived through
conventional methods. The quantitative RASC and CASC approach with
accompanying error analysis allows timing constraints t o be imposed on
basin modelling. In this way, the procedures allow important testing of
hypotheses. Figure 9.21 shows this procedure as applied t o Hibernia 0-35.
The burial history curve of Hibernia 0-35was derived using the program
BURSUB described in Gradstein et al. (1989). The minimum and
maximum age value as derived from error analysis of CASC was entered
 365

\\ A

I I I I I I I
1
180 150 120 90 60 30

AGE M a

Fig. 9.21 Burial history of Hibernia 0-35 well accounting for CASC derived error limits. Curve A has
minimum associated error, Curve B has maximum error.

as input into the program producing the two observed subsidence curves
shown. Such an approach provides an error envelope of burial curves
within which maturity calculations can be made which would help
determine the effect of chronology on the timing of peak generation and
expulsion of hydrocarbons.

Another application stems from an idea of Van Hinte (1984) who

describes t h e construction of s y n t h e t i c seismic sections from
biostratigraphic deata (the term synthetic isograrn is perhaps more
appropriate). The theory is quite simple and assumes that a seismic
section “is an image of time stratigraphic depositional patterns” (van
Hinte, 1984); and further: “...seismic section i s a record of
chronostratigraphic depositional and structural patterns and not a record
of time transgressive lithostratigraphy” (Vail and Mitchum, 1979).
Assuming that biostratigraphic correlation reflects these natural time
stratigraphic markers, the correlation of routinely produced CASC
isochrons (in 1 million year intervals) between suitable sections should
mimic seismic sections and allow reconstruction of the general geometry of
depositional sequences.

Such an isogram approach would not only enable a better integration

366

of seismic sections and paleontological data (i.e. t o determine if and where

the two do not key together) but will also allow the use of seismic
terminology of toplap, d o w n l a p , offlap a n d c o n c o r d a n c e w i t h
paleontologically derived schemes (with improved communication through
a common glossary of terms). Similarly, Van Hinte (1984) believes that
improved regional age calibrations of seismic sections would be apparent
as would be t h e i m p r o v e m e n t of t h e c o r r e l a t i o n of r e g i o n a l
seismostratigraphy to Vail and Mitchum’s (1979) Global Cycle Charts and
enhancing the understanding of the eustatics causing these changes i n
depositional styles (Van Hinte, 1984). The “isogram” between the wells
Flying Foam and West Flying Foam (Fig. 9.22) resembles seismic sections
between these wells and successfully predicted missing sections. Such
correspondence is testament to the predictive ability of the overall model.

9.10 Application of CASC to Palmer’s database

The application of RASC t o Palmer’s d a t a b a s e for t h e Riley

Formation in central Texas was discussed in Section 8.4. Table 8.13,
showed normality test output for the Morgan Creek, White Creek, and

1 WEST FLYING F O A M FLYING FOAM

2. -’

2.5 ..

E
x

X
3. -’

I-
P
Ly

3.5 -

Fig. 9.22 Isochron correlation between West Flying Foam and Flying Foam wells showing unconformity
and interpolations between known stratigraphic sections.
 367

Pontotoc sections. This information was combined with distances in feet

from base of section for each sample in order to create three DEP files for
use in CASC. RASC distance-depth curves were constructed with “depth”
measured in the stratigraphically downward direction taking new
reference points at 600,800 and 700 f t above base of section for the Morgan
Creek, White Creek and Pontotoc sections, respectively. This will permit
representation of CASC automated correlation lines together with
Palmer’s (1955) original zones and Shaw’s (1964) Riley Composite
Standard (R.S.T.) units for these three sections. The latter units were
constructed by Shaw in order to project his final composite standard back
onto each of the sections analyzed as a basis for biostratigraphic
correlation and isopach mapping of time intervals (Shaw, 1964, pp. 314-
316). The indirect method (called “cross-plots” in the CASC 2 module of
micro-RASC, see Chapter 10) was used for spline-curve fitting. Figure
9.23 shows the 30 RASC distances of Table 8.13 for the Morgan Creek
section plotted against the 18 event levels in this section. Table 9.12

RASC distance
6.0 4.0 2.0 0.0

‘ 20

Fig. 9.23 RASC distance-event level plot for Morgan Creek section. Spline-curve is for optimum (cross-
validation) smoothing factor SF = 0.382.
368

shows the results of cross-validation for selecting the best smoothing

factor.
The smoothing factor (SF = 0.38) derived from Table 9.12 is nearly
half as small as the standard deviation ( = 0.71) tacitly assumed in the
ordinary RASC method. This corroborates Shaw's assumption that the
Morgan Creek section has better t h a n average biostratigraphic
information. The optimum smoothing factor (SF = 0.37) for the White
Creek section which also is considered as better than average, is nearly the
same. (Shaw selected the Morgan Creek and White Creek sections as the
first two sections to be plotted against one another in the composite
standard method). Cross-validation applied t o the Pontotoc section,
considered to have the poorest biostratigraphic information by Shaw, gave
SF = 0.61.
Figure 9.23 shows the best-fitting spline-curve with SF = 0.38. It is
interesting t o compare this diagram with Figure 8.12 constructed for the
Morgan Creek section during application of modified RASC (see Section
8.9). The spline-curve of Figure 9.23 is more realistic. Although modified
RASC as a method has the capability of assigning different weights to
different stratigraphic events, it is not possible t o consider that the

TABLE 9.12

Smoothing factor (SF) and cross-validation value (CV) for RASC distance versus event level plot of
Morgan Creek section (Fig. 9.231. A. Minimum a n d maximum SF values correspond to f i r s t
monotonically increasing spline and best-fitting straight line, respectively. B. Zooming in on window
provided optimum value SF = 0.38.

0.364 0.2132 0.360 0.2187

0.392 0.2049 0.370 0.2076
0.420 0.2251 0.380 0.2023
0.448 0.2499 0.390 0.2041
0.476 0.2767 0.400 0.2094

0.504 0.3050 0.410 0.2171

0.532 0.3351 0.420 0.2245

0.560 0.3661 0.430 0.2333

0.588 0.3999 0.440 0.2425

0.616 0.4309 0.450 0.2519
0.435 0.4514 0.460 0.2612
 369

biostratigraphic information in some sections may be better than in

others. In this respect, CASC has the advantage because each section can
be analyzed separately in this technique and, on the average, the
deviations between points and fitted curves then will be less in sections
with better biostratigraphic information.

The curve of Figure 9.23 was combined with its line of observation
(depth-versus-level curve) to produce the RASC distance-depth plot of
Figure 9.24A. The lowest event (LO Kormugnostus simplex) with RASC
distance equal to 6.55 in the Morgan Creek Section is not shown in this
diagram which was redrafted from CASC 2 output. (The fitted curve does
not extend to 6.55 because some information was lost at the edges due t o
use of cross-plots). Figures 9.24B and C show similar plots for the White
Creek and Pontotoc sections. The standard deviations for the three curves
are equal to 0.39, 0.36 and 0.61, respectively, and nearly equal to the
optimum smoothing factors (see before). The three fitted curves become
steeper in the downward stratigraphic direction reflecting higher
sedimentation rate (cf. Shaw, 1964). Figure 9.24 can be used to determine
the probable depths of specific RASC distances in the three sections for
automated stratigraphic correlation. Figure 9.25 shows the results of the
CASC comparison together with Palmer’s zones and Shaw’s R.S.T. values.
The modified local error bars (k1 SD) shown for RASC distances 2.0, 5.0,
and 6.0 illustrate that the uncertainty increases in the downward
direction due to the higher sedimentation rate. The three sets of lines of
correlation agree closely with one another near the tops of the sections
where biostratigraphic control is relatively good. It is noted that the lines
for Palmer’s zones were drawn through the locations of the collections with
the highest stratigraphic position classified as belonging t o a particular
zone by Palmer (1955).

Shaw (1964) already had pointed out the good correspondence

between his R.S.T. values and Palmer’s zonation. In this section, it has
been shown that essentially the same results were obtained automatically
by using the RASCKASC approach. Although the highest occurrences of
individual fossils in the composite standard occur above the RASC
distances estimated for the same fossils, while the reverse holds true for
the lowest occurrences, this does not result in a systematic discrepancy
during correlation (cf. Chapter 8). In fact, when there is considerable
uncertainty in the biostratigraphic information (e.g. cuttings in wells
during exploration), it is preferable to base the lines of correlation on
average stratigraphic events instead of on subjective “total” ranges. This
370
RASC distance
6.0 4.0 2.0 0.0
I I I I I I 600

500 -
U

Y)

n
E
400
e
al

+
._
v)
n
300

RASC distance
6.0 4.0 2.0 0.0
I I I I I I I 800

-
-
I
LL
700
al
Y)

n
E
?
600
al

-5
C

v)

500

-
600 '
c

Y)

n
E
500
e
al
0
m
c
._
v)

Fig. 9.24 Spline-curves for positions of RASC distance values in three sections obtained by means of
indirect method. A. Morgan Creek section. Curve of Fig. 9.23 was combined with curve for positions of
event levels according to method of Fig. 9.6. Second (cf. Fig. 9.6b) and third (cf. Fig. 9.6e) smoothing
factors used were equal to 0.02 and 0.2, respectively Final standard deviation of deviations from curve is
SI)=0.390. B. White Creek section (SD=0.357). C. Pontotoc section (SD=0.615).
 371

Morgan Creek White Creek Pontotoc

Ft. Ft.

600 - 700

0.5 ).5
1.0
1.0 l.0
2.0 9.0
3.0
1.0

500 - 4.0 5.0 600

5.0 5.5

5.5

400 - 500
6.0
8.0

6.5
300 - 400
6.5

Fig. 9.25 Stratigraphic correlation of three sections by 3 methods. Palmer's (1955) zones and Shaw's
(1964) R.S.T. value correlation lines are superimposed on CASC results using spline-curves of Fig. 9.24.
Modified error bars extend one standard deviation on either side of probable positions for RASC distance
values equal to 2.0, 5.0 and 6.0, respectively. The uncertainty of the correlation lines increases in the
stratigraphically downward direction due to higher sedimentation rate.

is because, statistically speaking, the estimated average highest or lowest

occurrence for a taxon is closer t o its population value than the highest or
lowest occurrence on the range chart can be to its population value (cf.
Chapter 2).

9.1 1 Benthic foraminiferal zonation, central North Sea

The ranking and scaling method was used by Gradstein et al. (1988)
and Agterberg and Gradstein (1988) to propose a benthic foraminiferal
zonation for the Cenozoic deep marine deposits of the Central and Viking
372

Grabens, North Sea. Although CASC applications have not yet been
published, this case-history study is interesting because i t involves
integration of biostratigraphic and lithostratigraphic information, seismic
stratigraphy and correlation of Cenozoic hiatuses across the Atlantic
Ocean.
Following the widespread deposition of Danian chalk, south of about
60"N, the North Sea Basin underwent rapid subsidence (Sclater and
Christie, 1980; Gradstein and Berggren, 1981; Wood, 1981). As a result,
terrigenous clastic sediments in excess of 3 km thick accumulated in the

Fig. 9.26 Locations of 29 exploration wells, central North Sea

 373

central portion of the basin. Thickest sediments are found in the Central
Graben, whereas the Viking Graben received between 2 and 3 km of
sediment. Mudstones predominate, with deep marine clastic fans, like
those of the Forties and Frigg oil fields developing during the early stage
of Tertiary subsidence. In the Ekofisk area post-Danian olistostromes
occur. By Middle Miocene time the North Sea trough had been filled,
leaving a neritic environment with a predominantly calcareous benthic
microfauna dominated by Cassidulina, Elphidium, Fursenkoina, and
Cibicidoides.
The post-Danian, Paleocene through Early-Middle Miocene
mudstones harbour a rich and diversified flysch-type agglutinated benthic
fauna (Gradstein and Berggren, 1981), which includes over 60 taxa. Many
benthic taxa show minor and some major inconsistencies in relative
stratigraphic position of highest occurrence events as sampled in 29 wells
(Fig. 9.26). Over 2000 cuttings samples, sidewall cores, and some core
samples were analyzed, and the final analysis involves the tops of
147 benthics and relatively few planktonic taxa. The microfossil
distribution data were augmented by the relative positions in the wells of
physical log markers A through G as defined by A.C.Morton and
R.B. Knox (personal communication, 1984).
A close look at the North Sea analytical data shows that the southern
wells (blocks 21-38) contain more Oligocene-Miocene calcareous taxa,
including several species of planktonics, than the northern wells which
contain a more diversified Paleocene agglutinated record. This pattern of
geographic differentiation was further confirmed using correspondence
analysis (G.F. Bonham-Carter, personal communication, 1985). This
method clarifies the spatial distribution of co-occurring taxa. There may
be several reasons for this biogeographic trend, one of them being the fact
t h a t the principal deep water connection was t o the north in the
Norwegian Sea. The latter region does not have much of an indigenous
planktonic record. Another reason is that the post-Danian, Late
Paleocene-Eocene bathyal mudstone facies did not preserve much of a
carbonate record, owing to diagenetic effects (Gradstein and Berggren,
1981). A third reason is climatic; apparently the transition from carbonate
rich to carbonate poor rocks in Cenozoic time can be traced from south to
north over the central North Sea (Ziegler, 1981). The biogeographic
analysis indicates that for detailed regional studies two zonations are
required, one emphasizing the northern Paleogene record and the other
the southern Oliogocene-Miocene record. In this section, emphasis is on
374

the generalized zonation which combines features from both the Central
Graben and Viking Graben deep water troughs.
The generalized Cenozoic North Sea zonation uses the RASC
thresholds k , = 8 m,l = 1 and m,2 = 5, which means that zonal taxa must
occur in 8 or more out of 29 wells and each pair of taxa in the scaled
optimum sequence in 5 o r more wells. The threshold k , reduces the
original data set of 147 events t o 49 (Fig. 9.27), including 8 planktonic and
25agglutinated taxa and the log markers A-G of Knox and Morton as
found in the majority of the wells studied. The dendrograms that display
the interfossil distances between the ranked taxa (Fig. 9.27), are stable
when RASC is run with k , = 9 and 10 and m,2 = 6 and 7, which
incorporate 45 and 41 taxa, respectively. In each situation the same zones
are recognized.
In order to enhance the zonation with index taxa that are rare or
other taxa that are thought t o be potentially of such use, the RASC method
allows introduction of special or unique events (UE) occurring in one or a
few wells only. Twelve events were selected that occur in less than k , = 8
wells, but are worth noting. These events are the highest occurrences of
(from old to young) Ammodiscus planus, Reticulophragmium garcilassoi,
Bulimina trigonalis, Turrilina robertsi, Haplophragmoides (aff.) jaruisi,
Adercotryma sp. 1 (formally described as Adercotryma agterbergi, nsp. by
Gradstein and Kaminski, 19891, Globigerinatheka index, Turrilina
alsatica, Globigerina ex gr. officinalis, G. angustiurnbilicata and Neogene
radiolarian flood. In the final RASC calculations, stratigraphic neighbors
of these events are identified. A neighbor is a species that occurs in the
scaled optimum sequence and also in the wells with the UE, and
stratigraphically as close as possible to it. Each UE is positioned between
these neighboring events in the scaled optimum sequence (cf. Section 6.8).
Eleven interval zones are recognized (Figs. 9.27 and 9.28), with the
characteristic taxa listed stratigraphically in order of average

Fig. 9.27 Biozonation primarily based on agglutinated benthic foraminifers, Cenozoic, central North
Sea. The scaled optimum sequence is for the average tops of 54 foraminifers and siliceous microfossils
and physical log markers A-G in 29 wells. Dendrogram values a r e distances between events in relative
time. Scaling is stratigraphically downward, in line with the study of the wells. The generalized 10-
fold zonation is representative for the regional Cenozoic stratigraphy (see text). There a r e 11 unique
events ( = r a r e e v e n t s ) shown with * *, A s h a d i n g p a t t e r n h a s been used to e n h a n c e t h e
stratigraphically most useful parts of the dendrograms. The large interfossil distances a t the top of the
Danian, Late Selandian-Early Ypresian, Middle-Late Eocene, Late Oligocene-Early Miocene a n d
Middle Miocene a r e sedimentary cycle boundaries (from Agterberg and Gradstein, 1988).
3 75
376

disappearance. A more detailed zonation (which uses lower h, and mc2

values) is possible for local correlation. Gradstein et al. (1988) have shown
the approximate relation of the new zonation t o several standard
planktonic zones that can be recognized in the central North Sea, and to
the regional foraminifera1 zonation for the circum-central North Sea by
King (1983).

Approximately 30 taxa of agglutinated benthic Foraminifera have

distinct stratigraphic ranges in the central North Sea wells studied. A
comprehensive summary for the ranges is given in the range chart of
Agterberg and Gradstein (1988, pp. 21-26). So-called log markers are
valuable in the numerical stratigraphic analysis of the central North Sea.
These log markers are thought t o be chronostratigraphic in nature and
according to Morton and Knox (personal communication, 1984) correspond
t o the following approximate levels: marker G top Middle Miocene;

TIE POINTS FoRAMINIFEtlALZoNATloN

~ CHRONOSTWTI
AGE INTERPOLATION CENTRAL SEA
30 WELLS 71115 RUN
C ferefs -- QUATERNARY
_ _ _ _ ~ _ _
C sCaMisens!s ___II

G crassalormis

G praesc!lvla
1
r
. A gun&! (peak)

G praescilula

AQUITANIAN
G ollrcmals

T alsalica

LOG MARKER F
G mdex T pmeroli
R ampleclens BARTONIAN
R ampleclens
45
LUTETIAN $ G kuqlen

s pafagonrca
LOG MARKER D
LOG MARKER C
55 f 2P@=----'s. 0;
S pafagonca

S spclab~l~s
LCO

S pseudobulloides S tnloculino~des
S pseudobulloides
I I

Fig. 9.28 Relation between global model for (seismic) sequences stratigraphy (Vail, €fardenbol and
coworkers, pers. commun., 1986) and hiatuses based on scaling in time of the RASC zonations for the
central North Sea shown in Fig. 9.27 and the Canadian Atlantic margin shown in Fig. 6.2. Age tiepoints
for scaling are shown on each side of zonation in time. For explanation see text.
 377

marker F top Upper Eocene; marker E top Lower Eocene; marker D top
Sele Formation (or equivalent); marker C base Sele Formation (top
Paleocene); marker B top Ekofisk Formation (top Lower Paleocene); and
marker A top Cretaceous. The log picks were expected to vary slightly in
stratigraphic position relative to the foraminifera1 events in the wells and
were treated as “fossil events” in the calculations. Figure 9.27 shows the
calculated average stratigraphic position of these events. There is good
agreement between the ages assigned by Knox and Morton to the log picks
and the ages assigned t o the accompanying zones.
Log marker A is always found a t the level with Globotruncana below
the Danian zone (not shown in Fig. 9.27). Log marker B on average is in
the Danian, rather than at the top as suggested. Log markers C and D are
in the Coscinodiscus zone that delineates the ash-series that straddles the
Paleocene-Eocene boundary.
The top of log marker E is given as top of Lower Eocene in agreement
with its average occurrence slightly above the Subbotina patagonica zone,
Ypresian. The only serious exception to this average position was found to
be in well 23/22-1 where E occurs with Danian planktonics. The latter
may be reworked.

Log marker F occurs in the Globigerinatheka index interval, Upper

Eocene. Log marker G fits well at the top of the Globorotaliapraescitula-G.
zealandica zone, Lower-Middle Miocene. An interesting observation is
that the markers F and G are associated in Figure 9.27 with breaks in the
scaled optimum sequence. These breaks are recognized by large interfossil
distances between two events in adjacent zones. The large distance means
that there is little or no cross-over in position between the events in the
two successive zones. Such a situation is expected where there is a
stratigraphic section missing between the zones or a sudden change in
facies (may also be due to a hiatus). The latter is the case between the
Danian and Selandian zones, where carbonates are replaced by mudstones
and sands.
In computer runs without unique events such as Globigerinatheka
index, log marker F falls exactly at the large interfossil distance between
the Rotaliatina bulimoides zone and the underlying Reticulophragmium
amplectens zone. Log marker F marks the Eocene-Oligocene boundary in
the zonation and the large interfossil distance associated with it suggests a
hiatus involving the uppermost Eocene t o Lower Oligocene. In runs
378

including unique events, the Globigerinatheka index-log marker F events

are sandwiched between the R . bulimoides and R . amplectens zones,
indicating that these events are closely tied to the position of a hiatus in
many of the wells. Another break in the record is suggested a t the base of
the Globorotalia praescitula-G. zealandica zone, Oligocene-Miocene
boundary.
Log marker G reflects an Upper Miocene hiatus in the wells in the
central North Sea grabens. Few of the wells studied show unequivocal
fossil evidence for an upper Miocene interval (based on rare planktonic
foraminifers including Neogloboquadrina atlantica, left coiling and N .
acostaensis), and no RASC zone stands out.
In order to further investigate the stratigraphic extent of these breaks
or hiatuses in the central North Sea, a study was made of the most likely
sequence of events in Figure 9.27, but now scaled in linear time. First, 16
North Sea age tiepoints were assumed for a subset of fossil events in the
most likely (scaled) sequence with ages in million of years interpolated
from recent geochronological literature. Details on the age assignments
were given in Section 9.8 and in Gradsteinetal. (1988). Each of these
events also has a distance from the origin (top) in the scaled optimum
sequence. Next, the best fit was calculated between this series of events
scaled both in RASC units and in linear time using cubic spline fitting.
The resulting function of age versus distance may be used to convert the
RASC distances of all events in the scaled optimum sequence t o millions of
years. The result is shown in Table9.13, which also gives the original
tiepoints and their assumed ages. The spline-curve was only slightly
smoothed and passes nearly through all points with SF (Smoothing Factor)
of 0.37. As a result of this simple operation we can now stretch the North
Sea zones can be stretched in linear time and with the lower and upper
limits of the zones expressed in millions of years. This is shown in
Figure 9.28 (left) which also gives again the input ages of the tiepoints.
The scaling in linear time operation enhances the detection of breaks
or hiatuses that may occur in the central North Sea. Not unexpectedly,
only about 50% of Cenozoic time is represented by zones and their
representative sediments. For resolution and extent in time of the breaks
it is useful t o test this local time scale against similar ones based on other
fossil groups, like nannofossils and dinoflagellates, but such data is not
available at present. Individual error in age calibration does affect the
position of zonal boundaries, but not the general trend. On a local scale one
 379

TABLE 9.13

Interpolated ages of the events in the central North Sea zonation of Fig. 9.27, using cubic spline fitting
for the age-RASC distance relationship of a subset of events (shown a s *) for which age estimates (in
parentheses) are available in the literature.

Fossil RASC Interpolated Fossil RASC I ntrl-polatrd

distance age ( m y . ) distance age ( n 1 . v . )
31 * 0.3408 1.9(2) 117 6.0667 46.7
23 0.9841 4.2 68 6.1209 47 5
269 * 1.3056 6.1 (6) 264 6.1385 47.7
207 * 1.8527 10.9(11) 205 6.2 139 4x.x
219 2.0405 12.x 86 * 6.231 1 49.1 (49)
109 2.0548 12.9 260 6.2629 49.5
15 * 2.2503 14.9 (15) 50 * 6.55XS 53.4 ( S 5 )
236 * 2.5099 17.2(17) 263 6.6153 54.0
91 2.5529 17.6 54 6.6995 54.x
17 2.5561 17.6 45 7.0oxo 57.1
20 3.2268 22.7 279 7.1774 573
138 3.3777 23.7 277 7.3616 S7.Y
111 * 3.5783 24.9 (25) 204 * 7.4336 5X.1 ( 5 X )
25 3.7001 25.5 22 7.4353 5X.I
97 4.0920 27.7 136 7.5829 sx.3
I 82 4.1952 28.5 110 8.0943 59.1
142 4.4267 30.9 203 * X.lX37 59.2 (59)
140 * 4.5548 32.1 (31) I63 8.2012 59.2
24 4.5939 32.6 134 x.2~~2 50.4
1 83 4.7408 34.6 78 x.4590 59.7
262 4.8159 35.6 76 8.5015 SY x
206 * 4.8528 36.0 (37) 105 X.5021 S9.X
14X * 4.8713 36.2 (38) 57 * x.5921 h(1.U (60)
29 * 5.5634 40.7 (40) 65 R.6336 O(1.1
46 5.6672 41.6 I29 Y.0155 hi1 4
245 5.9375 44.9 25 I Y.UY?X hl I
26 I 6.0104 45.9 253 Y.7515 h.! X
61 * Y.X425 63.0 IhZ)

can expect improvements from more corroboration on the average

disappearance in time of the events used as tiepoints. For example, it is
assumed that R. bulimoides and T . alsatica disappear near 30 Ma, at the
end of the Rupelian, but this needs verification in more well sites.
Attention is drawn to the fact that the G. index extends over the Eocene-
Oligocene boundary, although G. index itself is Eocene. This taxon was
found reworked in Oligocene-Neogene deposits in several wells.

In order t o emphasize breaks of a more general nature, the

foraminifera1 zonation using RASC on the record in 27 Labrador and
Newfoundland offshore wells (see Fig. 6.27) was added to Figure 9.28, also
stretching it in linear time. Several age tiepoints, like S.pseudobulloides
380

(63 Ma), S.patagonica ( 5 5 Ma), T . robertsi (49 Ma). R . amplecterts (40 Ma),
and T . alsatica (30 Ma) are in common with the central North Sea
zonation. Again, large Eocene, Oligocene and Miocene hiatuses stand out.
Haq et al. (1987) have related a global seismic-sequence stratigraphy
to chronostratigraphy. The sequences are composed of periods of offlap
(basinward movement of the shoreline) and onlap (landwards movement of
the shoreline). These sequences are thought to reflect global changes in
sealevel. If rate of sealevel fall exceeds rate of basin subsidence, such
events can exert considerable influence on shallow deep marine clastic or
carbonate deposition. A relative shift seaward of the shoreline may
disrupt sedimentation in shallow basins, and lead to a hiatus. In deeper
water, more mass-flow sediments may occur causing local deposition or
erosion. The sequences were adjusted to conform to the linear time scale
used for the tiepoints (Berggren et al., 1985), and the North Sea and
Canadian Atlantic margin zones and the seismic sequence stratigraphy
were placed side by side (Fig. 9.28). Not unexpectedly, the more prominent
basinward shifts of the shoreline, for convenience numbered 1 through 7,
approximately coincide in time with breaks in the zonations. As discussed
earlier, large breaks in the scaled optimum sequence of fossil events are
likely to match hiatuses or sudden changes in facies. Major shifts in
position of shorelines influence the sediment supply as well as erosion and
can be expected t o exert control over the sedimentary sequences in the
Canadian offshore and Central and Viking Grabens. The latter, in turn,
influence the zonal boundaries of fossil assemblages.

Shift 1 in the North Sea may have coincided with replacement of the
Danian carbonates (S. pseudobuloides zone) by clastics ( R . paupera - T .
ruthuen murrayi zone). Shift 2 also is seen on the Rockall and Grand
Banks and may tie t o a late Ypresian hiatus. Shifts3 a n d 4 appear
associated with breaks in the uppermost Eocene and Oligocene, which
caused major disruptions in the fossil sequence both in Labrador and
North Sea wells. It is not easy to explain why in the deep central North
Sea a Late Eocene hiatus occurs. The mid-Oligocene shift 4 event appears
t o have affected the deeper North Sea less than the shallower beds offshore
Canada. This is t o be expected. Shift 5 does not match an Early Miocene
hiatus but events 6 and 7 bracket a Late Miocene break. In general, as
expected, the extent of the hiatuses and the presumed influence of sea
level changes increases stratigraphically upward with decreasing rate of
subsidence and sedimentation.
 38 1

Further study, particularly geological, will clarify the relation

between the global seismic-sequence stratigraphy and regional
sedimentary and paleontological history. It should be emphasized that the
RASC biochronology expresses average depositional sequence trends, not
necessarily reflected in each single well section. This brief case history on
the North Sea biostratigraphy and (local) biochronology has highlighted
the use of numerical methods t o advance the application of the fossil record
in subsurface geology.

i
600

1. Rut H-11 10.Lei1 E-38

2.Kerlsefni H-13 11. Freydis 8-87
3.Snorri J-90 12. Hare Bay H-31 ?
4. Herjoll M-92 13.Blue H.28
5.Bjarni H a l 14. Bonavista C-99
6. Gudrid H-55 15.Cumberland 8-55
7. Cartier D78 16. Bonanza M-71
8. Indian Harbour 17. Dominion 0.23
M-52 18. South Tempest
9, Lei1 M 4 G-88
55 0 19. Flying Foam 55"
1-13
20. Adolphua D-50
21. Hibernia P-15
22. Egret K-36
23. Osprey H-84
C A N A D A

ATLANTIC
500

OCEAN

6 450

- km 300

Fig. 9.29 Location map of the Labrador Shelf and Grand Banks wells used by DIorio and Agterberg
(1989).
382

9.12 Integration of foraminiferal and dinoflagellate datasets,

Labrador Shelf - Grand Banks
D’Iorio (1986, 1987, 1988) and D’Iorio and Agterberg (1989) have
dealt with the problem of using RASC and CASC for combined analysis of
microfossils belonging to different families. A RASC biozonation and
CASC correlation lines between 23 wells on the Labrador Shelf and Grand
Banks were based on the positions of highest occurrences of palynomorph
or foraminiferal taxa in wells and on their positions in a regional
biozonation model. The locations of these wells are shown in Figure 9.29.

The Cenozoic biozonation was established using the ranking and

scaling method (RASC). The automated correlation technique of CASC
provided a n effective method of identifying patterns of sediment
accumulation by tracing biozones through the wells of the study area. The

TABLE 9.14

Names and ages of biozones of Fig. 9.30 and list of boundary events used to trace RASC biozones in CASC
multi-well comparison.

Zone Age of Zone Name of Marker Event

I Paleocene Gavelinella beccaniformis
II Early Eocene Subbotina patagonica
Ill Early Middle Eocene Acannina densa
IV Late Middle Eocene Plectofrondiculana aff
paucicostata
V Late Eocene Reticulophragmium amplecfens
VI Late Eocene Turborotalia pomeroli
VII Oligocene Tumbna alsatm
Vlll Late Oligocene to
Early Miocene Uvigenna ex gr miozea nuttali
IX Middle Miocene Spiroplectamina cannata
X Middle to Late
Miocene Astengenna gc!nchi
XI Pliocene-Pleistocene Cassidulina feretis
_ _
Event
Boundary Number Event Name
I - II 52 Acannina soldadoensis
II - Ill 37 Acannina aff penfacamerata
I l l - IV 90 Acannina densa
IV - v 29 Reticulophragmium amplecfens
v - VI 263 Ammobaculifes aff polythalamus
VI - VII 259 Ammodiscus latus
VII - Vlll 24 Tumlina alsahca
Vlll - IX 21 Guttulina problem
IX - x 67 Scaphopodsp 1
x - XI 17 Astengenna gunchi
 383

CASC estimated depths of Cenozoic epoch boundaries agree well with

similar results determined from biostratigraphy. Slight differences
between the two estimates reflected systematic deviations a t the top and
bottom of the scaled optimum sequence. These trends were consistent with
results of a frequency distribution analysis of foraminifera1 last
occurrences by means of modified RASC. Histograms of the frequency
distributions of events were found t o be useful tools for identifying
potential marker events. These results could be used to trace vertical
migration of biostratigraphic events.
The construction and interpretation of the integrated biozonation
model has been described by D’Iorio (1987). The eleven zones identified
are named in Table 9.14. The discussion in this section will be restricted to
automated correlation (CASC multi-well comparison). The scaled
optimum sequence of Figure 9.30 was used as the correlation standard of
this study. Preliminary regional age estimates of events for CASC were
the same as those determined by D’Iorio (1987). The smoothing factors
were optimized with the cross-validation technique. The smoothing
factors which are equal t o the standard deviations of residuals used in the
CASC analysis of individual wells are listed in Table 9.15.

Labrador Shelf
Eleven wells were included in the Labrador Shelf group, the
southernmost one being Freydis. RASC biozones were correlated between
wells by tracing the depths of zone boundary events. These events were
chosen from Figure 9.30 and are listed in Table 9.14. When an event is not
found in a well, its expected depth was estimated from its RASC position.
The depths of the zone boundaries are listed in Table 9.15 and plotted in
Figure 9.31 (left side) for the Labrador Shelf wells.

The zone boundaries in the youngest or oldest parts of the wells may
not always be shown because of either the scarcity of data points, or the
specific shapes of the spline curves. The Bjarni, Cartier, Leif M-48 and
Freydis wells show more closely spaced zone boundaries, probably
indicating a lower sediment accumulation rate. This is in contrast with
the northern wells, which appear t o have greater sediment accumulation
rates.
384

Fig. 9.30 Biozonation model of the Cenozoic of the Labrador Shelf and Grand Banks based o n a n
integrated databank of foraminifers, dinoflagellates, and spores and pollen.
 385

Fig. 9.30(continued)

Zone boundaries reveal greater than average sediment accumulation

rates in zone IX in the Karlsefni and Herjolf wells, zone VIII in the Gudrid
well and zone VI in the Indian Harbour well. Low sediment accumulation
rates or an unconformity would explain the mutual proximity of zone
boundaries in Snorri from zones VI t o VIII, and in Freydis from zones I1
t o VII.
386

TABLE 9.15

CAW depths of biozone boundaries of Table 9.14. Errors are standard deviations.

Well Name -
I I1 11. 111 .
111 IV IV -v v - VI
Rut H-11 '3.16 f 0.63 '3.00 f 0.63
Karlselni H-13 371 ? 0 9 0 '2.94 t 0.20 '2.91 f 0.15 2 7 6 i 0 09 '2.74 f 0.10
Snorrt J-90 '2 9 5 f 0 0 6 '2 53 f 0 51 '2.48 f 0.21 2 15 f 0 05 2.14 t 0.03
Herloll M-92 '1 99 f 0 19 '1.97 i 0.18 176i021 1 72 i 0 22
Bjarni H-81 '1 99 f 0 12 '1 83 t 0 13 '1 82 f 0.12 1 6 7 1 0 08 '1 6 5 i 0 0 8
Gudrid H-55 '2 371.039 2 l o t 0 16 2.08t 0 16 19OfOll '1 88 i 0 09
Cartier D-70 '1.77 f 0.08 '1.76 f 0.08 1 55 i 0 26 1501017
Indian Harbour M-52 2 99 i 0 0 7 2.53 i 0.27 *
'2.47 0.35 2.28 f 0.32 216i047
Lei1 M-48 '1.69 f 0.07 '1.67 i 0.09 1 59 f 0.03 1.58 i 0.04
Lei1 E-38
Freydis 8-87 '1.54 i 0 10 'I 3 9 f 0 1 2 '1 38 f 0.11 1.28 i 0.05 '1 2 7 f 0.07
Hare Bay E-21 '3.06 i 0.12 '2.88 f 0.11 '2.86f0.10 2.27 f 0.67 2.18*0.26
Blue H-28 4.74 f 0.05 4.73 i 0.05 4.43 i 0.64 '4.29 ? 0.55
Bonavista C-99 '3.46 i 0.13 3.44 f 0.13 3.21 i 0.37 3.12f0.25
Cumberland B-55 3.58 f 0.18 '331 i 0 1 0 3291012 2.89i a.28 '2.81 t 0.28
Bonanza M-71 331i018 328i-018 2.90 i 0.40 2 78 i 0 63
Dominion 0-23 250r044 2 4 4 t 0 29 1.94 f 0.21 189*022
South Tempest G-88 230f-011 227f013 1.71 t o 2 5 1 59 f 0 44
Flying Foam 1-13 '1 9 6 i 0 1 5 *i.94ia.i8 1.67 f 0.16 1 62 f 0 23
Adolphus D-50 '2.63 f 0 10 2.28i 0.45 2.16 f 0.46 1.78 f 0.13 1 69 i 0 34
Hibernia P-15 1.25?0.19 120 i 0.10
Egret K-36
Osorev H-84 '0.76 i 0.06 '0.75 i 0 06

Well Name . VI VII

_______________~~~ ~
.
VII Vlll Vlll IX
~ IX -x x. XI
Rut H-11 2.24 t 0.08 '2 16 i 0.07 '1.74 i 0.52 '1 2 6 f 0.51 0.76 f 0.52
Karlselni H-13 '2.31 f 0.37 '2.15 t 0.08 '2.00 f 0.19 1 2 2 * 1.79 '0.69 f 0.11
Snorri J-90 '2.03 f 0.14 '1.83f0.12 1.73 i 0.04 1.50 t 0.68 '1.00 t 0.97
Herioll M-92 1.39 f 0.15 '1.29iO.12 '1.12iO.28 0.60 i 0.25
Bca;ni H-81 '1.42+0.19 1.21 f 0.34 0.87 i 0.23 0.64i 0.58
Gudrid H 55 178f006 1.72i0.11 0 70 f 0 35 '0 58 i 0 11 05iioia
Cartier D 70 131fO15 1.20 f 0.12 0 99 f 0 55 075i015 '0 66 f 0 09
Indian Harbour M 52 1 46 f 0 36 *
1 18 0.15 '0 96 f 0 42 '0 65 f 0 22 058f004
Lei1 M 48 '12OfO18 1.15f0.08 101+017 0 7 7 1 0 20 '0 5 6 f 0 36
Lei1 E-38 0.57 f 0.80 0.45 f 0 18 0.39 i 0 06
Freydts 6-87 '1.14t010 '0 87 f 0 51 +
0.67 0 41 0.51 f 0.19
Hare Bay E-21 1 56 f 0 22 1 35 t 0 25 '1 1 4 f 0 0 5 '1.08fO.10 '0.81 f 0 41
Blue H-28 '3 70 i 0 15 '3 52 f 0 21 *
'322 0 31 '2.96 f 0.29 '2.69 f 0.35
Bonavista C-99 2 37 f 0 15 1 95 f 0 88 121t o 4 9 '0.93 f 0.13 0.82f a.21
Cumberland 0-55 2 22 i 0 21 173i090 '1 1 0 i 0 3 7 '0.75f 0.16 0.63 f 0.16
Bonanza M.71 162fO.11 +
'1 52 0 09 1 39 i 0 08 '1.31 i 0 08 1 23 f 0.08
Dominion 0 23 1 30 t 0 33 1 05 t 0 29 0 *
.77 ~. 0 54 '0.57t 0 12 0 44 i 0.31
South Tempesl G-88 129iO17 117i013 '0 96 ?: 0 13 '0 82 f 0 05 0.76 f 0.06
Flying Foam 1-13 1 08 i 0 32 0 86 f 0 28 '0 61 i 0 21 '0 4 2 f 0 29 031i010
Adolphus D-50 122iO17 097i019 '0 54 i 0 15 '0 46* 0 03 '0 38 f 0 08
Hibernia P-15 0.99 i 0.05 0 8 1 1035 '0.43 i 0.29 '0 27 f 0 07
Egret K-36 *o 51 i a 05 0 48 i 0 03 0.43 i 0.08 '0 34 f 0 15
Osprey H-84 '0 62 i 0 12 '0 54 i 0 12 '0.43 tO.10 '0 37 f 0 08 0.32 i 0.09
~~

'The even1 used as a boundary indicator was not observed

Grand Banks
The Grand Banks group consists of twelve wells, the northernmost
one being Hare Bay. The Egret K-36 well is not included in the correlation
chart because it is shallow and has a very condensed section.

The zone boundary events listed in Table 9.14 also were traced in the
Grand Banks wells and plotted in Figure 9.31 (right side). The depth of
the boundaries and their respective local error estimates are presented in
Table 9.15.

A noticeable feature of Figure 9.31 (right side) is the thickening of

zone VI in the Bonanza well, suggesting a higher sediment accumulation
 387

0.0 1.0 2.0 3.0 4.0 5.0

0.0 1.0 2.0 3.0 4.0 5.0 + : : : : : : : : : : : : : : :
..........................

I I I I I I1 "In.,",. P-16 I

, : : : : : : : : : : : : : : : : : I - + : : . : : : : : + - ...........................
0.0 1.0 2.0 3.0 4.0 5.0 0.0 1.0 2.0 3.0 4.0 5.0

Depth (Km) Depth (Km)

Fig. 9.31 Biozone correlation chart of the Labrador Shelf wells (left side) and the Grand Banks wells
(right side). The zone boundaries are given in Table 9.14.

rate. The Osprey well exhibits relatively more closely spaced zone
boundaries than other wells; this is presumably due to its more distant
position from the terrigenous sediment supply (see Fig. 9.29). The Blue
well shows all zones at greater depths to the sea floor.

D'Iorio (1986) has shown that, for Cenozoic quantitative stratigraphy

of the Labrador Shelf - Grand Banks, combining different families of
microfossils resulted in biozonations of improved resolution. The results
described in this section illustrate that the RASC biozones can be traced
between all wells of the study area and that the CASC method effectively
identifies patterns of sediment accumulation rates. In D'Iorio and
Agterberg (1989) it also was shown that the CASC depth estimates of
Cenozoic age boundaries are consistent with the correponding boundary
depths assigned in the Atlas of the Labrador Sea (Srivastava, Editor,
1986).
This Page Intentionally Left Blank
 389

CHAPTER 10
COMPUTER PROGRAMS FOR RANKING, SCALING
AND REGIONAL CORRELATION OF STRATIGRAPHIC EVENTS

10.1 Introduction
The RASC computer program for r a n k i n g a n d s c a l i n g of
biostratigraphic events was originally written between 1978 and 1981 for
mainframe computers. It was followed by the CASC program for
correlation and scaling in time. In 1985, it became possible, after
relatively minor modification, to compile the FORTRAN code of the RASC
and CASC computer programs on IBM compatible microcomputers.
At present, several versions of these programs are in existence in
different languages (primarily FORTRAN, C and BASIC). A brief history
of the development of RASC and CASC with references is given a t the end
of this chapter. The existing programs are only slightly different from one
another. As a rule, later versions are more user-friendly than earlier ones.
The reader wishing to use RASC on a microcomputer (or mainframe) may
obtain a copy of Program RASC (Ranking and Scaling), version 12, which
at the time of writing (1990) is distributed free of charge by the Committee
on Quantitative Stratigraphy (CQS). (Please send 360 KB floppy diskette
to F.M. Gradstein, Chairman, CQS, Atlantic Geoscience Centre, Bedford
Institute of Oceanography, Dartmouth, N.S., Canada, B2Y 4A2). This
enhanced batch version of RASC in FORTRAN 77 by Agterberg et al.
(1989) contains source code, executable (EXE) files and test data files. It
can be executed on a PC with math co-processor. CASC is available as a
mainframe program (Agterberg et al., 1985). Agterberg and Byron (1990)
are preparing micro-RASC for release as a Geological Survey of Canada
Open File.
The micro-RASC system consists of 12 separate program modules. It
makes use of the characteristic features of microcomputers. Except for
Module 1 which can be used to create new input files, each module reads
one or more input files and creates one or more output files. This allows
flexibility for program development because separate modules can be
revised and replaced without changing the remainder of the system.
390

Micro-RASC contains slightly modified code previously published in

the RASC (RAnking and Scaling) and CASC (Correlation And Scaling in
time) computer programs for r a n k i n g , scaling a n d correlation of
biostratigraphic events. This code has been supplemented by more
recently developed algorithms including use of the jackknife method for
estimating variances of cumulative RASC distances in scaling, modified
RASC for frequency distribution analysis of stratigraphic events, and
cross-validation to decide on optimum smoothing factors i n spline-curve
fitting for CASC. Micro-RASC can be used on any IBM compatible
microcomputer with math co-processor and a FORTRAN compiler. This
version of RASC and CASC is exclusively numeric using simple graphics
programmed in FORTRAN 77.

The contents of the 12 modules is summarized in the next section with

references to sections in earlier chapters where more details can be found.
Important decisions to be made by the RASC user in order to create the
parameter (PAR) file needed for r u n n i n g t h e modules a r e listed
separately, i n Section 10.3, together with those parameters t h a t can be
changed from their default values. These decisions a r e numbered as in
micro-RASC. They are of a general nature in that any RASC user should
consider the questions asked here. Parts of Module 1 (Data input) were
previously published by Heller et al. (1985). Modules 2, 3,4,5 and 6 are
equivalent to RASC version 12 (Agterberg et al., 1989). Modules 7 and 8
can be added to RASC version 12 relatively quickly, by using t h e
procedures described in Chapter 7 (Jackknife scaling) a n d Chapter 8
(Modified RASC). An earlier version of Module 8 consisting of FORTRAN
and BASIC programs was included in D’Iorio (1988). Earlier versions of
Modules 9 and 11 were distributed on a n informal basis under the names
TSREG (Regional time scales) a n d SPLIN (Spline-curve f i t t i n g ) ,
respectively. Modules 10 and 12 emulate mainframe CASC.

It should be kept in mind that the purpose of the RASC computer

programs is t o order and correlate stratigraphic events. I n most
applications, t h e events a r e stratigraphically h i g h e s t a n d lowest
occurrences of (micro-)fossils, although peak occurrences and a b r u p t
changes in relative abundance can be used equally well for correlation if
these can be defined systematically. Lithostratigraphic, seismic and
magnetostratigraphic events can be combined with biostratigraphic
events. However, these other types of events may need special
consideration; e.g. by defining them as marker horizons or by evaluating
their relative uncertainty independently by means of modified RASC for
 391

frequency distribution analysis. Although the RASC computer programs

provide automatic stratigraphic correlations, the user should remain i n
control of input and output, e.g. the input can be modified or amplified on
the basis of new information provided in successive outputs.

10.2 S u m m a r y of contents of the 12 modules of micro-RASC

Module 1: DATA INPUT (cf. Section 4.2)

The RASC method requires as input a sequence (SEQ) file with coded
sequences of stratigraphic events for individual sections, a dictionary with
event names (DIC file), and a parameter (PAR) file with settings of
switches and values of parameters. The CASC method requires depth
(DEP) files for individual sections. Module 1 allows preparation of d a t a
(DAT) files from which SEQ files and preliminary DEP files are generated
automatically. Examples of DAT file formats are:

(a) Depths (in feet or metres) followed by dictionary code numbers of

events (feet will be converted into meters); and

(b) Fossil code numbers followed by depths of lowest a n d highest

occurrences (DIC file for use in RASC then will be created with
separate entries for lowest and highest occurrence of each fossil).

Module 2: PREPROCESSING (cf. Section 4.8)

Frequencies of events a r e determined as follows: ( a ) number of

sections for each event; (b) number of events occurring in h sections and
number of events occurring i n h or more sections ( h = l , 2, ..., n; n
represents total number of sections). The threshold parameter h, must be
selected. Further analysis will be restricted to events that occur in at least
h, sections. Special rare events (“unique” events) which a r e to be re-
inserted later in the biozonation, but which occur in fewer than h, sections,
should be identified. It is also possible to define marker horizons (e.g.
seismic e v e n t s or b e n t o n i t e l a y e r s ) w h i c h a r e n o t s u b j e c t t o
biostratigraphic uncertainty.
392

Module 3: RANKING (cf. Sections 5.3 and 5.5)

The optimum sequence of events is determined by probabilistic
ranking or “presorting” with or without the modified Hay method. The
sequence obtained by presorting may be improved by sorting on the basis
of superpositional relations (“above”, “below”, coeval) between pairs of
events using the modified Hay method. Inconsistencies involving three or
more events (cycles) will be identified (cf. Sections 5.7 and 5.8). The
threshold parameter rn,l may be selected by changing its default value
rn,l = 1. The modified Hay method will be applied only t o pairs of events
occurring in at least rn,l sections. It may not be possible to determine the
relative order of two or more events. This type of uncertainty is expressed
by means of the uncertainty range (Section 5.4) assigned t o all events in
the optimum sequence.

Module 4: SCALING (cf. Section 6.3)

The scaled optimum sequence of events is determined by estimating
intervals between successive events in the optimum sequence previously
obtained by ranking. This process usually involves minor reordering of
the events. Final distances between successive events are clustered in a
dendrogram which is useful as a regional biozonation. The threshold
parameter rnzc must be selected. Scaling calculations are restricted t o
frequencies of order relations between pairs of events occurring in at least
m,2 sections. Unweighted and weighted scaling can be performed.
Further analysis (e.g. normality test) will be based on weighted scaling
with the weights determined by frequencies of superpositional relations
between events. Standard deviations of inter-event distances are
provided. The cumulative RASC distance of each event is computed and
added t o the preliminary DEP files in order to create (complete) DEP files
if CASC will be used.

Module 5: RANK EVALUATION (cf. Section 7.3)

The optimum sequence resulting from Module 3 or 4 can be used for
construction of the occurrence table, “step model” and scattergrams.
Events are shown for individual sections in the occurrence table. Their
observed sequence in each section is compared with the optimum sequence
 393

in the step model. Penalty points are assigned for each position that an
event is out of place in a section. Kendall’s rank correlation coefficient
can be computed from the total number of penalty points per section. The
relative order of events in each section is compared to that in the optimum
sequence in the scattergrams.

Module 6: NORMALITY TEST (cf. Sections 6.6 and 8.2)

The observed sequence of events in each section is compared to the
scaled optimum sequence using cumulative RASC distances. Second-order
differences are computed for events by comparing their observed positions
to those of their neighbors in the stratigraphically upward and downward
directions. Events that are out of place with probabilities greater than
95% and 99% are identified. A frequency distribution analysis of second-
order differences is performed in order to evaluate: (a) autocorrelation of
successive events in the scaled optimum sequence; and (b) overall
frequencies of anomalous events occurring either too low (e.g. due t o
contamination during drilling), o r too high (e.g. due t o geological
reworking) in the sections.

Module 7: JACKKNIFE SCALING (cf. Section 7.5)

In scaling, each estimated distance ag between successive events (i
and j ) is the average of a number of primary distance estimates Dij,k based
on the superpositional relations of i and j with other events ( h ) . By
successively deleting individual events, and scaling of the reduced data-
sets, it is possible t o obtain measures of precision of the estimates by
means of the jackknife method which is non-parametric. This also results
in jackknife estimates of the cumulative RASC distances of the events.
Only if the latter estimates are close to the original cumulative RASC
distances, as obtained by Module 4,their jackknife standard deviations
can be used as standard deviations of the cumulative RASC distances.

Module 8: MODIFIED RASC (cf. Section 8.5)

The scaling method is based on transforming frequencies for
superpositional relations between events into fractiles of the normal
394

distribution in standard form. I t is assumed that all events have the same
variance for deviations between their regional mean positions a n d
observed positions within individual sections. In modified RASC, t h e
variances of the events can be different. They are estimated by means of
a n iterative procedure. Firstly, spline-curves are fitted to the events i n
common between the scaled optimum sequence and individual sections in
order t o project the regional mean positions onto the sections, and to collect
all deviations for each event. Secondly, the variance of the deviations for
each event is used for scaling which yields a new set of cumulative RASC
distances. These two steps are repeated until approximate convergence is
reached. Modified RASC allows identification of low-variance events
which can be used a s marker horizons. In addition t o different event
variances, this procedure provides frequency distributions of individual
events which may be positively o r negatively skewed. Maximum
deviations can be used for constructing a conservative range chart i n
which the ranges are based on regional highest and lowest observed
occurrences of fossils (cf. Sections 8.7 to 8.9).

Module 9: REGIONAL TIME SCALE (cf. Section 9.8)

The age (in millions of years) may be known for a subgroup of the
stratigraphic events used in a regional RASC study. I t may be possible to
establish the relationship between age and cumulative RASC distance for
this subgroup. This relationship, expressed a s a spline-curve, can be used
to transform all RASC distances into ages in linear time. These ages can
be used to replace the cumulative RASC distances in the DEP files for
CASC. Events can be weighted differently, either by using standard
deviations resulting from Module 7, or by using subjective weights.

Module 10: CASC 1: EVENT-DEPTH CURVES (cf. Section 9.3)

The CASC (Correlation And Scaling in time) method consists of two
steps: (a) construction of event-depth curves (this module or Module 11);
and (b) multiwell comparison (Module 12). Module 10 closely resembles
the age-depth curve-fitting p a r t of the CASC mainframe computer
program. Supplementary statistical techniques (cross-validation,
jackknife spline-curve fitting) are given in Module 11 which amplifies
Module 10. Input for CASC in Module 10 or 11consists of DEP files for the
 395

sections t o be studied. A spline-curve is fitted for each section. The

dependent variable is (a) rank, (b) RASC distance, or (c) age in Ma; the
independent variable is (a) relative event level, or (b) depth (in metric
units). Because events generally are spaced irregularly along the depth-
axis, the indirect method can be used for estimating the age-depth curve.
This algorithm called “cross-plots” consists of fitting separate spline-
curves for the age-level and depth-level relations, Elimination of level
then gives a n event-depth curve which usually is better t h a n the one
obtained by direct spline-curve fitting. Sediment a c c u m u l a t i o n
(sedimentation) rate curves can be obtained from the first derivatives of
the spline-curves.

Module 11: CASC 2: STATISTICAL ANALYSIS (cf. Section 9.4)

The shape of a spline-curve is to a large extent controlled by its
smoothing factor (SF) representing the s t a n d a r d deviation of t h e
differences between observed and fitted values. The law of superposition
of strata requires that age never decreases i n the stratigraphically
downward direction. This provides a n e s t i m a t e of t h e m i n i m u m
smoothing factor. The maximum smoothing factor correspond to the best-
fitting straight line of least squares. The optimum smoothing factor has a
value within the open interval bounded by these two extremes. Cross-
validation is a method for estimating the optimum smoothing factor. The
best-fitting spline-curve deviates from the unknown true age-depth curve.
The error of the fitted spline-curve values can be estimated by using the
jackknife method. The method for spline-curve fitting in Modules 10 and
11 is a modified version of De Boor’s (1978) FORTRAN program. Module
12 allows two alternative techniques for spline-curve fitting. These are
discrete cubic spline fitting (Duris, 1980) a n d a beam deformation
analogue method (Hibbert, 1990), respectively.

Module 12: CASC 3: MULTI-WELL COMPARISON (cf. Section 9.8)

Probable depths of selected events or isochrons (e.g. multiples of
10Ma) determined by means of the age-depth curves can be correlated
between sections. This multi-well comparison is performed by means of a
table in which the probable depths are accompanied by estimated 68% or
95% confidence intervals. Various types of confidence intervals can be
396

obtained. These include the local and modified local error bars for
deviations between observed depth of events and the probable depths used
for correlation. Local and modified local error bars basically are error bars
along the time axis which have been projected along the depth axis by
assuming locally constant and variable rates of sediment accumulation,
respectively.

10.3List of decisions to be made by user of the RASC computer

programs
During a complete micro-RASC session, the user can be asked 80
questions numbered separately in each of the twelve modules. The answer
to each question is “yes” or “no’’. If the answer is “yes”, the switch
corresponding to the question is turned on. It is left off if the answer is
“no” and a default decision would be made which is displayed on the
monitor. The user then is given the chance t o change “no” into “yes”.
Some questions are asked only if certain conditions are satisfied. Eleven
questions are about a parameter with a default value t h a t can be
changed. The settings of the switches and the values of the parameters are
entered in the PAR file needed to run the micro-RASC programs. At the
beginning of each module, the user is asked if the switches are to be set for
that module. If the answer is “no”, an existing PAR file must be used.

Module 1: DATA INPUT

1.1 Do you wish to prepare a new dictionary?
Default: It will be assumed that you work with an existing dictionary.

1.2 Do entries represent stratigraphic events?

Default: It will be assumed that you wish to work with the highest and lowest occurrences of
fossils.

1.3 Do you wish to make a HI and LO occurrences dictionary?

Default: It will be assumed that a HI and LO dictionary is in existence, and will not have to be
created from a single entry dictionary.

1.4 Are you working in the stratigraphically downward direction?

Default: It will be assumed that you work in the stratigraphically upward direction.
 397

1.5 Will you work with the depths of the samples?

Default: It will be assumed that you work with event levels along a relative depth scale.

1.6 Do you wish to enter rotary table height and water depth?
Condition: Switch 1.5 ison.

1.7 Are your depths metric?

Condition: Switch 1.5 is on.
Remark: If Switch 1.7 is turned on, the following supplementary question is asked: Are your
depths in meters? If the answer to the supplementary question is “no”, the user is asked to: Enter
conversion factor from meters to the units of your depth (Example: if your depths a r e in
kilometers, enter 1000).
Default: It is assumed that you work with feet. These will be automatically changed into metric
units.

1.8 Do you want the depth files for use in CASC?

Default: It will be assumed that you will not wish to use CASC.

1.9 Do you wish to create preliminary depth files?

Default: I t will be assumed that your depth files already exist.

1.10 Do you wish to create a new sequence (RASC input) file?

Default: It will be assumed that you work with an existing sequence file.

1.11 Do you wish to create a new data file?

Default: It will be assumed that you work with a n existing data file.

1.12Do you wish to subtract a constant from all dictionary numbers that
are read in?
Parameter name: NSTART (Default value NSTART = 0).
Default: As usual, no changes are made in the dictionary numbers.

Module 2: PREPROCESSING
2.1 Do you wish to set the threshold parameter for minimum number of
sections in which a n event should occur?
Parameter name: IOCR (Default value: IOCR = 3)
Default: The minimum number of sections in which a n event should occur is equal to 3.
398

2.2 Are you dealing with two separate groups of fossils which should have
different threshold parameters?
Condition: Switch 1.12 ison.
Parameter name: IOCR2 (Default value: IOCR2 = 0)
Default: As usual, you wish to use a single threshold parameter for minimum number of sections.

2.3 Do you wish to define unique events? (i.e. special rare events that
occur fewer than IOCR times)
2.4 Do you wish to define marker horizons?
2.5 Do you wish to see intermediate tabulations?
Default Intermediate tabulations (e.g. recoded sequence data) will not be shown in the output.

Module 3: RANKING
3.1 Do you wish t o perform presorting?
3.2 Do you wish to apply the modified Hay method?
3.3 Do you wish t o set the threshold parameter for minimum number of
sections in which a pair of events should occur?
Parameter name: CRITl (Default value: CRITl = 1.0)
Default: All frequencies will be used for the modified Hay method.

3.4 Do you wish to re-set the tolerance?

Parameter name: TOL (Default value: TOL=O.O)
Default As usual, the tolerance parameter is kept equal to zero.

3.5 Do you wish t o change the maximum number of iterations?

Parameter name: ITER (Default value: ITER = 10,000)
Default: The maximum number of iterations allowed for the modified Hay method is 10,000.

3.6 Do you wish t o see the cycling tabulations?

Default: The cycling tabulations will not be shown in the output.

3.7 Do you wish to see all intermediate tabulations?

Default Intermediate tabulations (e.g., matrices with initial and reordered frequency scores) will
not be shown in the output.

3.8 Do you wish t o go on t o the scaling module?

Default: RASC run will be terminated after ranking and input for Module 4 will not be created.
 399

3.9 Do you wish to perform ranking evaluation?

Default: Input for ranking evaluation (Module 5) will not be created.

3.10Do you wish to add ranking results to depth files for use in CASC?
Condition: Switch 1.9 is on.
Default: As usual, CASC will not be applied to ranking results.

3.11 Do you wish to re-insert unique events into the optimum sequence?
Default: Unique events will not be re-inserted into the optimum sequence.

Module 4: SCALING
4.1 Do you wish to set the threshold parameter for minimum number of
sections in which a pair of events should occur?
Parameter name: CRITP (Default value: CRIT2 = 2.0)
Default: All frequencies for pairs occurring in two or more sections will be used for scaling.

4.2 Do you wish to change the truncation limit?

Parameter name: AAA (Default value: AAA = 0.95)
Default: Frequency of a n event observed to occur above another event in all sections (containing
both events) will be changed from 1.00 to 0.95.

4.3 Do you wish to delete scaling tables from output?

Default: Only dendrograms will be shown in the output.

4.4 Should long distances be suppressed during estimation?

Default: As usual, long distances will not be suppressed.

4.5 Should final reordering be applied?

4.6 Do you wish to apply scaling more than five times before accepting
the final reordering results?
Condition: Switch 4.5 is on.
Parameter name: KKL (Default value: KKL = 5)
Default: Total number of interations during reordering is not allowed to exceed 5.

4.7 Do you wish to see intermediate tabulations?

Default: Intermediate tabulations (tables of fractiles) will not be shown in the output.
400

4.8 Do you wish to suppress re-insertion of unique events into the scaled
optimum sequence?
Default: As usual, unique events will be re-inserted into the scaled optimum sequence.

4.9 Do you wish to perform rank evaluation?

Default: Rank evaluation (Module 5 ) of scaling results will not be performed.

4.10 Do you wish to perform the normality test?

Default: Normality test (Module 6) will not be performed.

4.11 Do you wish to perform jackknife scaling?

Default: Jackknife scaling (Module 7)will not be performed.

4.12 Do you wish to apply the modified RASC method?

Default: Modified RASC (Module 8) will not be performed.

4.13 Are you planning to construct a regional time scale using ages (in Ma)
of selected events?
Default Regional time scale (Module 9) will not be constructed.

4.14Do you wish to add scaling results to depth files for use in CASC?
Condition: Switch 1.9 is on; Switch 3.10 is off.

Module 5: RANK EVALUATION

5.1 Do you wish to construct the occurrence table?
5.2 Do you wish to apply the step model?
5.3 Do you wish to see scattergrams for separate sections?

Module 6: NORMALITY TEST

6.1 Do you wish to see the detailed statistical analysis results? 6.e. study
of autocorrelation based on second-order differences).

Module 7: JACKKNIFE SCALING

7.1 Do you wish to change the width of the window on the RASC scale?
P a r a m e t e r name: WDW (Default value: WDW = 2.0)
 40 1

Default: No use will be made of observed superposional relations between events that are above
one another in the original scaled optimum sequence with a probability of approximately 95
percent.

7.2 Do you wish to use the jackknife standard deviations for construction
of a regional time scale?
Condition: Switch 4.13 is on.

Module 8: MODIFIED RASC

8.1 Do you wish to perform more than three complete iterations?
Parameter name: KKM (Default value: KKM = 3)
Default: As usual, the cumulative RASC distance estimates will be refined three times using
successive approximations of the event variances.

8.2 Do you wish to see frequency tables for separate events?

8.3 Do you wish to see plots of observed and calculated values for separate
sections?
8.4 Do you wish to construct the range chart table?
8.5 Do you wish to save the event variances for weighting in CASC?
Condition: Switch 4.14 is on.

Module 9: REGIONAL TIME SCALE

9.1 Do you want to use automated version?
Condition: Switch 7.2 is on.
Default You will choose your own smoothing factor for spline-curve fitting with age a s the
dependent variable.

9.2 Do you wish to define subjective weights in order to assign more or

less influence to ages of events?
Default: All ages will have equal weights during spline-curve smoothing.

9.3 Do you want to substitute ages for RASC distances in depth files?
Condition: Switch 4.14 is on.
Default: CASC will be based on the RASC distances.
402
Module 10: CASC 1: EVENT-DEPTH CURVES
10.1 Are you using an optimum sequence with ranks only?
Condition: Switch 3.10 or Switch 4.14 is on.
Default: You are using the scaled optimum sequence supplemented by RASC distances or ages (in
Ma).

10.2If some events are observed to be coeval, do you wish to work with
separate events at approximately the same event levels?
Default: Events observed to be coeval a t a given level will be averaged.

10.3 Should each average for a n event level be weighted according to the
numbers of coeval events on which it is based?
Condition: Switch 10.1is off.

10.4Do your depth files contain standard deviations for separate events
which are not equal t o one another?
Condition: Switch 10.1is off.
Default: All events will be weighted equally.

10.5 Are you using weights determined by means of modified RASC?

Condition: Switches 8.4and 10.3 are on; Switch 10.1 is off.

10.6 Will you be performing a multi-well comparison?

Default: Age-depth results will not be saved for multi-well comparison.

10.7 Will you use the indirect method for estimating event-depth
relations?
Default: The direct method will be used for estimation.

10.8Do you want to study the first derivatives and sediment accumulation
curves?
10.9 Do you wish to use defaults except for the age-level relation?
Condition: Switch 10.7 is on.
Default: You will have to select smoothing factors for the event-depth and age-interpolated depth
relations in each section.
 403

10.10Do you wish to use the minimum smoothing factor and other defaults
in all sections?
Condition: Switch 10.6 is on; Switch 10.7 is off.
Default: Sections will be analyzed separately one after another.

10.11Do you wish to use plot axes defined during analysis of the first depth
file later, for the other depth files?
Default: You can let the program define default plot axes or define new plot axes for any section.

10.12Do you wish to perform detailed statistical analysis (e.g. cross-

validation) for at least one of your sections?
Default: I t will not be possible to use Module 11.
R e m a r k If Switch 10.12 is off, the next prompt asks for the name of the first depth file to be
analyzed by means of Module 10.

Module 11: CASC 2: STATISTICAL ANALYSIS

11.1 D o you wish to use cross-validation?
Condition: Switch 10.12 is on and Module 11 has been activated.
Default: Optimum smoothingfactor will be determined by autocorrelation method.

11.2 D o you wish to see additional tabulations (e.g. spline coefficients) in

the output?
11.3 D o you wish t o obtain the jackknife spline-curve?
11.4 D o you wish t o use discrete cubic spline smoothing?
Default: As usual, a modification of De Boor’s program for cubic spline smoothing will be used.

11.5 D o you wish to use the beam deformation analogue method for cubic
spline smoothing?
Condition: Switch 10.2 is on;Switch 10.4 is off.
Default: As usual, a modification of De Boor’s program for cubic spline smoothing will be used.
R e m a r k The next prompt asks for the name of the first depth file to be analyzed by means of
Module 11.
404
Module 12: CASC 3: MULTI-WELL COMPARISON
12.1 Do you wish to specify the sections to be used for correlation?
Default: All sections analyzed by means of Module 10 or Module 11 will be used for correlation.

12.2 Do you wish to correlate selected events?

Default: Your correlation will be based on ages in millions of years.

12.3 Do you wish to compute probable positions of isochrons?

Condition: Switch 12.2 isoff.
Default: Ages for correlation between sections will have to be selected individually

12.4 Do you want modified local error bars?

Default: Local error bars will be given only.

12.5 Do you want approximate 95 per cent confidence intervals?

Default: Standard deviations will be used for the error bars (i.e., approximate 68 per cent
confidence intervals will be given).

12.6 Do you wish to define a new t-value for the error bars?
Condition: Switch 12.5 is on.
P a r a m e t e r name: TVALUE (Default value: TVALUE = 2.0)
Default: As usual, the approximation t = 2.0 for 95 per cent confidence intervals will be used

12.7 Do you want statistical analysis results for spline-curve values and
studentized residuals as well?
Default: As usual, statistical analysis will be restricted to deviations between observed and
calculated values.

10.4 Brief history of the development of RASC and CASC

The basic ideas incorporated in the RASC RAnking and Scaling

computer program originated during 1978 in collaboration with F.M.
Gradstein (Bedford Institute of Oceanography in Dartmouth, Nova Scotia,
Canada). For initial program development in FORTRAN IV, use was
made of the Cyber 74 computer of the Department for Energy, Mines and
Resources in Ottawa. Agterberg and Nel(1982a,b) published the ranking
and scaling algorithms in the journal “Computers & Geosciences”.
Stratigraphic and statistical model verification with applications i n
exploration biostratigraphy in petroleum basins were given in Gradstein
and Agterberg (1982).
 405

During spring, 1979, a n earlier version of the program was

implemented by W.A. Burroughs on the DECSystem 10 of Syracuse
University and tested by graduate students participating in a seminar on
quantitative stratigraphic correlation. Their comments and discussions
with J.C. Brower (Syracuse University) resulted in many improvements.
The program also was implemented by K.G. Shih and A. Johnston a t the
Bedford Institute of Oceanography for demonstration in August, 1979,
during the first meeting of t h e Canadian Working Group of the
International Geological Correlation Programme (IGCP) Project 148
(Quantitative Stratigraphic Correlation Techniques). Suggestions were
received during this workshop and later from participants including
P.H. Doeven (PetroCanada, Calgary, Canada), L.E. Edwards (U.S.
Geological Survey, Reston, Virginia, U.S.A.), P. Moore (Shell Resources
Canada, Calgary, Canada), E.M. Oliver (Robertson Research, Calgary,
Canada), and R.J. Price (Amoco Canada, Calgary, Canada). A number of
results obtained by RASC were presented during the second meeting of the
Canadian working group for ICGP Project 148 in Ottawa, February, 1980
(Agterberg and Gradstein, 1981). This included comprehensive scaling
studies carried out in Ottawa by C.B. Hudson (University of South
Carolina, Columbia, U.S.A.; see Hudson and Agterberg, 1982), and
presentation of RASC output using DISSPLA by A. Jackson (Bedford
Institute of Oceanography, Dartmouth, Canada). The version of RASC
published in “Computers & Geosciences”was implemented by S. Briggs on
the DECSystem 10 and IBM 370 computers of Syracuse University during
spring, 1981.
An interactive version of mainframe RASC using a Tektronix 4014
terminal was prepared with the help of C.F. Chung (Geological Survey of
Canada, Ottawa) and R. Lessard (University of Sherbrooke, Quebec) and
used for demonstration during the Second International Quantitative
Stratigraphy Short Course held during the Calgary 1982 meetings of the
American Association of Petroleum Geologists (co-sponsored by IGCP
Project 148, the Canadian Society of Petroleum Geologists and the
University of Calgary). New implementations by oil companies including
use on the UNIVAC 1108 by Shell Resources Canada in Calgary resulted
in further suggestions for improvement. M. Heller and W.S. Gradstein
(Consultants in Halifax, Nova Scotia) prepared a user guide for RASC
which was released in 1983 as Geological Survey of Canada (GSC) Open
File 922 (Heller et al., 1983). This Open File also contained revised
406

FORTRAN IV code for mainframe RASC (printout with examples and

magnetic tape).

I n 1982, with t h e help of J. Oliver (University of O t t a w a ) ,

development of CASC (Correlation And Scaling in time) commenced in
Ottawa (Agterberg and Gradstein, 1983). This interactive program was
developed in FORTRAN IV using a Cyber 730 mainframe with Tektronix
4014 terminal. The program was demonstrated d u r i n g t h e Third
I n t e r n a t i o n a l Q u a n t i t a t i v e S t r a t i g r a p h y S h o r t Course, held i n
Dartmouth, Nova Scotia, October 1983 and the Seventh Meeting of the
Canadian Working Group for IGCP Project 148 held in Ottawa, March
1984. The CASC program was released in 1985 as GSC Open File 1179
(Agterberg et al., 1985). Applications of CASC are described in Gradstein
and Agterberg (1985), Williamson (1987) and D’Iorio and Agterberg
(1989).

By 1985, microcomputer hardware and software had advanced to the

stage that RASC could be run on IBM PC’s and compatibles equipped with
the 8087 Math Co-processor. S.N. Lew prepared a FORTRAN 77 version of
RASC which, together with the revised user’s manual, was published as
GSC Open File 1203 (Heller et al., 1985). This program can be compiled
and run on microcomputers. GSC Open Files 1179 and 1203 can be
obtained from the Publications Office of the Geological Survey of Canada,
601 Booth Street, Ottawa K1A OE8 (Each Open File consists of a manual
and two 5.25-inch double-sided, double-density diskettes with IBM-PC
readable code; cost $20.00 for OF 1179 and $25.00 for OF 1203). The
DEN0 program (Jackson et al., 1984) serves to display dendrograms of
scaled optimum sequences and the optimum sequences of stratigraphic
events from RASC output by means of a CALCOMP plotter. It is written
in the plotting language DISSPLA.

Alethic Software Incorporated (52 Parkhill Road, Halifax, Nova

Scotia, B3P 1R5) has developed three computer programs in the language
C for IBM personal computers (XT, AT, PS2) and compatibles (with math
co-processor). Their GEOSCI-1 program is for data entry. It prepares
sequence files and dictionaries for GEOSCI-2 which is a C version of
RASC. Alethic’s GEOSCI-3 program is a C version of CASC. These
programs are marketed by Alethic and can be obtained a t the address
shown above or by phoning 902-423-9860.
 407

Assisted by D. Gillis (Atlantic Geoscience Centre, Dartmouth, N.S.),

F.M. Gradstein introduced output redirection to the code of RASC for IBM-
PC. This feature improves its use on microcomputers t h a t usually lack
high-speed printers. This version (RASC 011) was later used during the
Eighth International Quantitative Stratigraphy Shortcourse held at the
Free University, Amsterdam, February 1989.

With the help of S.N. Lew (Geological Survey of Canada, Ottawa), a

FORTRAN 77 program called SPLIN for microcomputers using IBM
Graphics Development Toolkit was developed for spline-fitting of age-
depth curves with cross-validation. Use was made of De Boor’s (1978)
FORTRAN programs. SPLIN w a s demonstrated d u r i n g the F i f t h
International Quantitative Stratigraphy Short Course held in Aberdeen,
Scotland, April 1986. For method and applications, see Agterberg and
Gradstein (1988)and Gradstein et al. (1989). Discussions with M. Fearon
(Consultant in Halifax, Nova Scotia) resulted i n improvements of the
spline-fitting algorithm.

SPLIN was combined with a microcomputer version of CASC with the

help of J. Kirk (Informatics Applications Division, Energy, Mines and
Resources Canada, Ottawa). This program (SPLIN2) was demonstrated
during the Seventh International Quantitative Stratigraphy Short Course
held at PETROBRAS, Rio de Janeiro, Brazil, November 1987. Modified
RASC for frequency distribution analysis of biostratigraphic events
(Agterberg and D’Iorio, in press) was developed in collaboration with M.A.
D’Iorio (1988) whose doctoral dissertation contains FORTRAN 77 and
BASIC programs for an earlier version of modified RASC.

Development of the micro-RASC computer programs was commenced

in 1988 with the help of D. Byron a t the Geological Survey of Canada in
Ottawa, with contributions by P. Hibbert (Informatics Applications
Division, Energy, Mines and Resources Canada, Ottawa). F.M. Gradstein
provided valuable contributions by reviewing the blueprints several times
with many comments. Micro-RASC will be available a s a GSC Open File
(Agterberg and Byron, 1990a,b).

During 1989, Z. Huang (Dalhousie University) added “help” files and

enhanced output formats of the batch version of RASC (equivalent to
Modules 2 to 6 of micro-RASC). This program (RASC, version 12) is
available through t h e Committee on Quantitative Stratigraphy (see
Section 10.1 on how to obtain it).
This Page Intentionally Left Blank
 409
REFERENCES
Agterberg, F.P., 1974. Geomathematics. Elsevier, Amsterdam, 596 pp.
Agterberg, F.P., 1984. Binomial and trinomial methods in quantitative biostratigraphy. Computers and
Geosciences, 10: 31-41.
Agterberg, F.P. (Editor), 1984. Theory, Application and Comparison of Stratigraphic Correlation
Methods. Comput. Geosci., 10 (1):1-186.
Agterberg, F.P., 1985. Normality testing and comparison of RASC to Unitary Associations Method. In:
F.M. Gradstein et al., Quantitative Stratigraphy, UNESCO, Paris and Reidel, Dordrecht,
pp. 243-262.
Agterberg, F.P., 1988. Quality of time scales - a statistical appraisal. In: D.F. Merriam (Editor), Current
Trends in Geomathematics, Plenum, New York, pp. 57-103.
Agterberg, F.P. and Bonham-Carter, G.F. (Editors), 1990. Statistical applications in the Earth Sciences.
Geological Survey of Canada Paper 89-9.
Agterberg, F.P. and Byron, D.N., 1990a. FORTRAN 77 microcomputer programs for ranking, scaling
and regional correlation of stratigraphic events. In: F.P. Agterberg and G.F. Bonham-Carter
(Editors), Statistical Applications in the Earth Sciences, Geol. Surv. Can. Paper 89-9.
Agterberg, F.P. and Byron, D.N., 1990b. Micro-RASC System of 12 FORTRAN 77 microcomputer
programs for ranking, scaling and regional correlation of biostratigraphic events. Geological
Survey of Canada Open File.
Agterberg, F.P. and DIorio, M.A., in press. Frequency distributions of highest occurrences of Cenozoic
Foraminifera along the northwestern Atlantic Margin. Proceedings, 4th South American
COGEODATA Symposium, held in Our0 Preto, Brazil, November 1987.
Agterberg, F.P. and Gradstein, F.M., 1981. Workshop on quantitative stratigraphic correlation. Math.
Geol., 13(1):81-91.
Agterberg, F.P. and Gradstein, F.M., 1983. Interactive system of computer programs for stratigraphic
correlation. Current Research, Geological Survey of Canada, Paper 83-1A,pp. 83-87.
Agterberg, F.P., and Gradstein, F.M., 1988,Recent developments in quantitative stratigraphy. Earth-
Science Reviews, 25(1): 1-73.
Agterberg, F.P. and Nel, L.D., 1982a. Algorithms for the ranking of stratigraphic events. Computers
and Geosciences, 8: 69-90.
Agterberg, F.P. and Nel, L.D., 1982b. Algorithms for the scaling of stratigraphic events. Computers and
Geosciences, 8: 163-189.
Agterberg, F.P. and Rao, S.N. (Editors), 1988. Recent Advances in Stratigraphic Correlation. Hindustan
Publishing Corporation, Delhi, 192 pp.
Agterberg, F.P., Gradstein, F.M., Lew, S.N. and Thomas, F.C., 1985. Nine databases with applications of
ranking and scaling of stratigraphic events. In: F.M. Gradstein et al., Quantitative Stratigraphy,
UNESCO, Paris and Reidel, Dordrecht, pp. 473-564.
Agterberg, F.P., Gradstein, F.M. and Nazli, K.,1990. Correlation of Jurassic microfossil abundance data
from the Tojeira sections, Portugal. In: F.P. Agterberg and G.F. Bonham-Carter (Editors),
Statistical Applications in the Earth Sciences, Geol. Surv. Can. Paper 89-9.
Agterberg, F.P., Gradstein, F.M., Nel, L.D., Lew, S.N.,Heller, M., Gradstein, W.S., D'Iorio, M.A.,
Gillis, D. and Huang, Z.,1989. Program RASC (Ranking and Scaling) version 12. Comm.
Quantitative Stratigraphy, Bedford Inst. Oceanogr., Dartmouth, N.S., Canada.
Agterberg, F.P., Oliver, J., Lew, S.N., Gradstein, F.M. and Williamson, M.A., 1985.CASC Fortran IV
interactive computer program for correlation and scaling in time of biostratigraphic events.
Geological Survey of Canada Open File Report 1179.
Armstrong, R.L., 1978. Pre-Cenozoic Phanerozoic time scale. In: G.V.Cohee, M.G.Glaessner and
H.D. Hedberg (Editors), Contributions to the Geologic Time Scale, Am. Assoc. Petroleum Geol.,
Studies in Geology, No. 6,pp. 73-91.
Barrell, J., 1917. Rhythms and the measurements of geologic time. Geol. SOC. Am., Bull., 28: 745-904.
Baumgartner, P.O., 1984. A Middle Jurassic-Early Cretaceous low-latitude radiolarian zonation based
on Unitary Associations and age of Tethyan radiolarians. Eclogae Helv., 71: 729-837.
Baumgartner, P.O., 1987. Age and genesis of Tethyan Jurassic radiolarites. Eclogae Geol. Helv., 30:
831-879.
Berge, C., 1973. Graphes et Hypergraphes. Dunod, Paris, 516 pp.
Berger, W.H. and Heath, G.R., 1968. Vertical mixing in pelagic sediments. J. Marine Res., 26: 134-143.
Berggren, W.A., 1972. A Cenozoic time-scale, some implications for regional geology and
paleogeography. Lethaia, 5: 195-215.
410
Berggren, W.A., Kent, D.V., Flynn, J.J. and Van Couvering, J.A., 1985. Cenozoic geochronology. Geol.
SOC.Am. Bull., 96: 1402-1418.
Blank, R.G., 1979. Applications of probabilistic biostratigraphy to chronostratigraphy. J . Geol., 87:
647-670.
Blank, R.G., 1984. Comparison of two binomial models in probabilistic biostratigraphy. Computers and
Geosciences, 10: 59-67.
Blank, R.G. and Ellis, C.H., 1982. The probable range concept applied to the biostratigraphy of marine
microfossils. J. Geol., 90: 415-433.
Bliss, C.I.,1935. The calculation of the dosage-mortality curve. Ann. Appl. Biol., 2 2 134-167.
Blow, W.H., 1969. Late middle Eocene to Recent planktonic foraminifera1 biostratigraphy. In:
P. Bronnimann and J.H.H. Renz (Editors), Proc. 1st International Conf. on Planktonic
Microfossils, Geneva 1967, E.J. Brill, Leiden, pp. 339-378.
Bonham-Carter, G.F., Gradstein, F.M. and DIorio, M.A., 1986. Distribution of Cenozoic Foraminifera
from the Northwestern Atlantic Margin analyzed by correspondence analysis. Computers and
Geosciences, 12: 621-635.
Box, G.E.P. and Jenkins, G.M., 1976. Time Series Analysis: Forecasting and Control. Holden-Day, San
Francisco, 575 pp.
Bramlette, M.N. and Sullivan, F.R., 1961. Coccolithophorids and related nannoplankton of the Early
Tertiary in California. Micropal., 7: 129-188.
Brinkmann, R., 1929. Statistisch-biostratigraphische Untersuchungen a n Mitteljurassischen
Ammoniten: Uber Artbegriff und Stammesentwicklung. Abhandlungen der Gesellschaft der
Wissenschaften zu Cattingen, Mathematisch-Physikalische Klasse, Neue Folge 13(3), 249 pp.
Brower, J.C., 1981. Quantitative biostratigraphy, 1830-1980. In: D.F. Merriam (Editor), Computer
Applications in the Earth Sciences, Plenum, New York, pp. 63-103.
Brower, J.C., 1985a. Multivariate analysis of assemblage zones. In: F.M. Gradstein et al., Quantitative
Stratigraphy, UNESCO, Paris and Reidel, Dordrecht, pp. 65-94
Brower, J.C., 1985b. Archaeological seriation of a n original data matrix. In: F.M. Gradstein e t al.,
Quantitative Stratigraphy, UNESCO, Paris and Reidel, Dordrecht, pp. 95-108.
Brower, J.C., 1990. A case study for comparison of some biostratigraphic techniques using Paleogene
alveolinids from Slovenia and Istria. In: F.P. Agterberg and G.F. Bonham-Carter (Editors),
Statistical Applications in the Earth Sciences, Geol. Surv. Can. Paper 89-9.
Brower, J.C., Millendorf, S.A. and Dyman, T.S., 1978. Quantification of assemblage zones based on
multivariate analysis of weighted and unweighted data. Computers and Geosciences, 4 221-227.
Brunk, H.D., 1960. Mathematical models for ranking from paired comparisons. J . Am. Stat. Assoc., 55:
503-520.
Burroughs, W.A. and Brower, J . C . , 1982. SER, a FORTRAN program for t h e s e r i a t i o n of
biostratigraphic data. Computers and Geosciences, 8: 137-148.
Buzas, M.A., Koch, C.F., Culver, S.J. and Sohl, N.F., 1982. On the distribution of species occurrence.
Paleobiology, 8: 143-150.
Carinati, R., Marini, A. and Potenza, R.G., 1982. The mathematical formalization of the geological
relations identifying the basic structure of a geological relations identifying the basic structure of
a geological data bank. In: J.M. Cubitt and R.A. Reyment (Editors), Quantitative Stratigraphic
Correlation, Wiley, Chichester, pp. 13-18.
Carr, P.F., Jones, B.G., Quinn, B.G. and Wright, A.J., 1984. Toward a n objective Phanerozoic time scale.
Geology, 12: 274-277.
Car&, B., 1979. Graphs and Networks. Clarendon Press, Oxford, 277 pp.
Cheetham, A.H. and Deboo, P.B., 1963. A numerical index for biostratigraphic zonation in the mid-
Tertiary of the Eastern Gulf. Gulf Coast Association of Geological Societies, Transactions, 13:
139-147.
Christopher, R.A., 1978. Quantitative palynologic correlation of three Campanian and Maestrichtian
sections (Upper Cretaceous) from the Atlantic coastal plain. Palynology, 2 1-27.
Clark, R.M., 1989. A randomization test for the comparison of ordered sequences. Math. Geol., 21:
429-442.
Cowie, J.W. and Bassett, M.G.(Compilers), 1989. International Union of Geological Sciences, 1989
Global Stratigraphicchart. Supplement to Episodes, 12(2).
Cox, A.V. and Dalrymple, G.B., 1967. Statistical analysis of geomagnetic reversal data and the precision
of potassium-argon dating. J. Geophysical Res., 72: 2603-2614.
Craven, P. and Wahba, G., 1979. Smoothind noisy data with spline functions. Numerische Mathematik,
31: 377-403.
 411

Cross, T.A. (Editor), 1990. Quantitative Dynamic Stratigraphy. Prentice Hall, Englewood Cliffs, New
Jersey, 625 pp.
Cubitt, J.C. (Editor), 1978. Quantitative Stratigraphic Correlation. Comput. Geosci., 4 (3): 215-318.
Cubitt, J.C. and Reyment, R.A. (Editors), 1982. Quantitative Stratigraphic Correlation. Wiley,
Chichester, U.K.,320pp.
Davaud, E., 1982. The automation of biochronological correlation. In: J.M. Cubitt and R.A. Reyment
(Editors), Quantitative Stratigraphic Correlation, Wiley, Chichester, pp. 85-99.
Davaud, E. and Guex, J., 1978. Traitement analytique ‘manuel’ et algorithmique de problhmes
complexes de correlations biochronologiques. Eclogae Geol. Helv., 71: 581-610.
David, H.A., 1988. The Method of Paired Comparisons (Second Edition). Oxford Univ. Press, New York,
N.Y., 200 pp.
David, M., 1977. Geostatistical Ore Reserve Estimation. Elsevier, Amsterdam, 364 pp.
Davidson, R.R., 1970. On extending the Bradley-Terry model to accommodate ties in paired comparison
experiments. J . Amer. Statist. Assoc., 65: 317-328.
Davis, J.C., 1986. Statistics and Data Analysis in Geology, 2nd Edition. Wiley, New York, N.Y., 646 pp.
De Boor, C., 1978. A Practical Guide to Splines. Springer Verlag, New York, 392 pp.
Dienes, I., 1974. General formulation of the correlation problem and its solution in two special
situations. Math. Geol, 6: 73-81.
Dienes, I., 1982. Formalized Eocene stratigraphy of Dorog Basin, Transdanubia, Hungary, and related
areas. In: J.M. Cubitt and R.A. Reyment (Editors), Quantitative Stratigraphic Correlation,
Wiley, Chichester, pp. 19-42.
Dienes, I. and Mann, C.J., 1977. Mathematical formalization of stratigraphic terminology. Math. Geol.,
9: 587-603.
D’Iorio, M.A., 1986. Integration of foraminifera1 and dinoflagellate d a t a sets in quantitative
stratigraphy of the Grand Banks and Labrador Shelf. Bull. Canadian Petroleum Geology, 34:
277-283.
D’Iorio, M.A., 1987. Quantitative biostratigraphic analysis of the Cenozoic of 23 Canadian Atlantic
offshore wells. The Compass, 64: 264-277.
DIorio, M.A., 1988. Quantitative biostratigraphic analysis of the Cenozoic of the Labrador Shelf and
Grand Banks: Unpublished Ph.D. thesis, Univ. of Ottawa, 404 p.
DIorio, MA., 1990. Sensitivity of the RASC model to its critical probit value. In: F.P. Agterberg and
G.F. Bonham-Carter (Editors), Statistical Applications in the Earth Sciences, Geol. Surv. Can.
Paper 89-9.
D’Iorio, M.A. and Agterberg, F.P., 1989. Marker event identification technique and correlation of
Cenozoic biozones on the Labrador Shelf and Grand Banks. Bull. Canadian Petroleum Geol., 37:
346-357.
Dixon, W.J. and Massey, F.J.,1957. Introduction to Statistical Analysis. McGraw-Hill, New York, N.Y.,
488 pp.
Doeven, P.H., 1983. Cretaceous nannofossil stratigraphy and paleoecology of the Canadian Atlantic
Margin. Bull. Geol. Surv. Can. no. 356,70 pp.
Doeven, P.H., Gradstein, F.M., Jackson, A,, Agterberg, F.P. and Nel, L.D., 1982. A quantitative
nannofossil range chart. Micropal., 28: 85-92.
Doveton, J.H., 1986. Log analysis of Subsurface Geology Concepts and Computer Methods. Wiley, New
York, N.Y., 273 p.
Drobne, K., 1977. Alveolines Pal6oghnes de la Slovhie et de 1’Istrie. MBm. Suisses Paleont., 99,175 pp.
Drooger, C.W., 1974. The boundaries and limits of stratigraphy. Proc. Kon. Ned. Akad. Wet. Ser. l l B ,
17: 159-176.
Duris, C.S., 1980. Algorithm 547, FORTRAN routines for discrete cubic spline interpolation and
smoothing. ACM Transact. Math. Softw., 6: 92-103.
Edwards, L.E., 1978. Range charts and no-space graphs. Computers and Geosc., 4: 247-258.
Edwards, L.E., 1982. Numerical and semi-objective biostratigraphy: Review and predictions. Proc. 3rd
North Am. Pal. Conv., Montreal, August 1982,l: 147-152.
Edwards, L.E., 1984. Insights on why graphic correlation (Shaw’s method) works. J . Geology, 92:
583-597.
Edwards, L.E., 1989. Supplemented graphic correlation: A powerful tool for paleontologists and
nonpaleontologists. Soc. Econ. Paleontologists and Mineralogists, Research Reports, pp. 127.143.
Edwards, L.E. and Beaver, R.J., 1978. The use of paired comparison models in orders stratigraphic
events. J. Math. Geol. 10: 261-272.
Efron, B., 1982. The Jackknife, the Bootstrap and Other Resampling Plans. SOC.for Industrial and
Applied Mathematics, Philadelphia, Pennsylvania, 92 pp.
412

Eubank, R.L., 1984. The hat matrix for smoothingsplines. Statist. and Prob. Letters, 2: 9-14.
Eubank, R.L., 1988. Spline Smoothing and Nonparametric Regression. Dekker, New York, N.Y.,
438 pp.
Finney, D.J., 1971. Probit Analysis (3rd Edition). Cambridge Univ. Press, 333 pp.
Fisher, R.A. and Yates, F., 1964. Statistical Tables for Biological, Agricultural and Medical Research
(6th Edition). Oliver and Boyd, Edinburgh, 146 pp.
Foster, N.H., 1966. Stratigraphic leak. Am. Assoc. Pet. Geol. Bull., 50: 2604-2606.
Fulkerson, D.R. and Gross, O.A., 1965. Incidence matrices and interval graphs. Pacific J. Math. 15:
835-855.
Gale, N.H., Beckinsale, R.D. and Wadge, A.J., 1980. Discussion of a paper by McKerrow, Lambert and
Chamberlainon the Ordovician, Silurian and Devonian time scales. Earth Plan. Sc. L., 51: 9-17.
Gill, D. and Merriam, D.F. (Editors), 1979. Geomathematical and Petrophysical Studies in
Sedimentology. Pergamon, Oxford, 266 pp.
Gilmore, P.C. and Hoffman, A.J., 1964. A characterization of comparability graphs and interval graphs.
Can. J. Math. 6: 539-548.
Glenn, W.A. and David, H.A., 1960. Ties in paired-comparison experiments using a modified
Thurtone-Mosteller model. Biometrics, 16: 86-109.
Golub, G.H., Heath, M. and Wahba, G., 1979. Generalized cross-validation as a method for choosing a
good ridge parameter. Technometrics, 21: 215-223.
Gordon, A.D., 1982. An investigation of two sequence-comparison statistics. Austral. J . Statistics, 24:
332-342.
Gordon, A.D., Clark, A.M. and Thomson, R., 1988. The use of constraints in sequence slotting. In:
E. Diday (Editor), Data Analysis and Informatics V, North Holland Publishing Co., Amsterdam,
pp. 353-364.
Gradstein, F.M., 1984. On stratigraphic normality. Computers and Geosciences, 10: 43-57.
Gradstein, F.M., 1985. Ranking and scaling in exploration micropaleontology. In: F.M. Gradstein et al.,
Quantitative Stratigraphy, Unesco, Paris and Reidel, Dordrecht, pp. 109-160.
Gradstein, F.M. and Agterberg, F.P., 1982. Models of Cenozoic foraminiferal stratigraphy -
Northwestern Atlantic Margin. In: J.M. Cubitt, and R.A. Reyment (Editors), Quantitative
Stratigraphic Correlation, Wiley, Chichester, pp. 119-173.
Gradstein, F.M., and Agterberg, F.P., 1985. Quantitative correlation in exploration micropaleontology.
In: F.M. Gradstein et al., Quantitative Stratigraphy, UNESCO, Paris and Reidel, Dordrecht,
pp. 309-360.
Gradstein, F.M. and Berggren, W.A., 1981. Flysch-type agglutinated foraminifera and t h e
Maestrichtian to Paleogene history of the Labrador and North Seas. Marine Micropal., 6: 211-268.
Gradstein, F.M. and Fearon, M., 1990. STRATCOR, a new method for biozonation and correlation with
applications to exploration micropaleontology (Summary). In: F.P. Agterberg and G.F. Bonham-
Carter (Editors), Statistical Applications in the Earth Sciences, Geol. Surv. Pap. 89-9.
Gradstein, F.M. and Kaminski, M.A., 1989. Taxonomy and biostratigraphy of new and emended species
of Cenozoic deep-water agglutinated Foraminifera from the Labrador and North Seas. Micropal.
35: 72-92.
Gradstein, F.M. and Srivastava, S.P., 1980. Aspects of Cenozoic stratigraphy and paleogeography of the
Labrador Sea and Baffin Bay. Palaeogeogr., Palaeoclimatol., Palaeoecol., 30: 261-295.
Gradstein, F.M. and Williams, G.L., 1976. Biostratigraphy of the Labrador Shelf, I. Geol. Surv. Canada
Rept. 349.40 pp.
Gradstein, F.M., Agterberg, F.P., Aubry, M.-P., Berggren, W.A., Flynn, J.J., Hewitt, R., Kent, D.V.,
Klitgord, K.D., Miller, K.G., Obradovitch, J . , Ogg, J.G., Prothero, D.R. and Westerman, G.E.G.,
1988. Sea level history. Science 241: 599-605.
Gradstein, F.M., Agterberg, F.P., Brower, J . C . and Schwarzacher, W.S., 1985. Quantitative
Stratigraphy. Unesco, Paris, and Reidel, Dordrecht, 598 p.
Gradstein, F.M., Agterberg, F.P. and D'Iorio, M.A., 1990. Time in quantitative stratigraphy: In:
T.A. Cross (Editor),Quantitative Dynamic Stratigraphy. Prentice-Hall, Englewood Cliffs, N.J.,
pp. 519-542.
Gradstein, F.M., Fearon, J . M . and Huang, Z.,1989. BURSUB and DEPOR version 3.50 -Two FORTRAN
77 programs for porosity and subsidence analysis. Geol. Surv. Can. Open File 1283.
Gradstein, F.M., Kaminsky, M . and Berggren, W.A., 1988. Cenozoic foraminiferal stratigraphy of the
Central North Sea. In: F. Rogl and F.M. Gradstein (Editors), Proc. 2nd Agglutinated Foraminifera
Workshop, Vienna, 1986, Abhandlungen der Geologischen Bundesanstalt, 41: 97-108.
 413
Gradstein, F.M., Williams, G.L., Jenkins, W.A.M. and Ascoli, P., 1975. Mesozoic and Cenozoic
stratigraphy of the Altantic continental margin, eastern Canada. In: G.T. Yorath et al. (Editors),
Canada's Continents Margin and Offshore Petroleum Exploration, Can. Soc. Petroleum Geol.
Mem. 4, pp. 103-121.
Grimm, E.C., 1987. CONISS: A FORTRAN 77 program for stratigraphically constrained cluster
analysis by the method of incremental sum of squares. Computers and Geosciences, 13: 13-35.
Guex, J., 1977. Une nouvelle mbthode d'analyse biochronologique, note prbliminaire. Bull., Soc. Vaud.
Sci. Nat., 73: 309-321.
Guex, J., 1980. Calcul, caractbrisation et identification des associations unitaires en biochronologie.
Bull. SOC.Vaud. Sci. Nat., 75: 111-126.
h e x , J., 1981. Associations virtuelles et discontinuit& dans la distribution des esp4ces fossiles: un
exemple inthressant. Bull., Soe. Vaud. Sci. Nat., 75: 179-197.
Guex, J., 1987. Corrdations biochronologiques et Associations unitaires. Presses Polytechniques
Romandes, Lausanne, Switzerland, 264 pp.
Guex, J., 1988. Utilisation des horizons maximaux rbsiduels en biochronologie. Bull., Soc. Vaud. Sci.
Nat., 79.2: 135-142.
Guex, J. and Davaud, E., 1984. Unitary associations method: Use of graph theory and computer
algorithm. Computers and Geosciences, 10: 69-96.
Gyji, R.A. and McDowell, F.W., 1970. Potassium argon ages of glauconites from a biochronologically
dated Upper Jurassic sequence of northern Switzerland. Eclogae Geol. Helvetiae, 63: 11-118.
Hald, A., 1957. Statistical Theory with Engineering Application. Wiley, New York, N.Y., 783 pp.
Hald, A, 1960. Statistical Tables and Formulas. Wiley, New York, 97 pp.
Hallam, A., 1975. Jurassic Environments. Cambridge Univ. Press, Cambridge, 269 pp.
Haq, D.U., Hardenbol, J . and Vail, T.R., 1987. Chronology of fluctuating sea levels since the Triassic.
Science, 235: 1156-1166.
Hardenbol, J.,Vail, P.R. and Ferrer, J., 1981. Interpreting paleoenvironments; subsidence history and
sea-level changes of passive margins from seismics and biostratigraphy. Oceanologica Acta 1981,
sp., pp. 33-44.
Harland, W.B., Cox, A.V., Llewellyn, Pickton, C.A.G., Smith, A.G. and Walters, R., 1982. A Geologic
Time Scale. Cambridge Univ. Press, 131 pp.
Harper, C.W., Jr., 1981. Inferring succession of fossils in time: The need for a quantitative and
statistical approach. J. Paleont., 55: 442-452.
Harper,C.W., Jr., 1984. A Fortran IV program for comparing ranking algorithms in quantitative
biostratigraphy. Computers and Geosciences, 10: 3-29.
Hay, W.W., 1972. Probabilistic stratigraphy. Eclogae Geol. Helv., 65: 255-266.
Hay, W.W. and Southam, J.R., 1978. Quantifying biostratigraphic correlation. Annual Review of Earth
and Planet Sc., 6: 353-375.
Hazel, J.E., 1977. Use of certain multivariate and other techniques in assemblage zonal biostratigraphy,
examples utilizing Cambrian, Cretaceous, and Tertiary benthic invertebrates. In: E.G. Kauffman
and J.E. Hazel (Editors), Concepts and Methods and Biostratigraphy, Dowden, Hutchinson and
Ross, Stroudsburg, Pennsylvania, pp. 187-212.
Hedberg, H.D. (Editor), 1976. International Stratigraphic Guide. Wiley, New York, N.Y., 200 pp.
Heller, M., Gradstein, W.S., Gradstein, F.M. and Agterberg, F.P., 1983. RASC FORTRAN IV computer
program for ranking and scaling of biostratigraphic events. Geological Survey of Canada Open
File 922.
Heller, M., Gradstein, W.S., Gradstein, F.M., Agterberg, F.P. and Lew, S.N., 1985. RASC Fortran 77
computer program for ranking and scaling of biostratigraphic events. Geological Survey of
Canada Open File 1203.
Hemelrijk, J., 1952. A theorem on the sign test when ties are present. Kon. Nederl. Akad. Wetensch.,
Proc., 55. 322.
Hibbert, P., 1990. Spline smoothing by means of an analogy to structural beams. In: F.P. Agterberg and
G.F.Bonham-Carter (Editors), Statistical Applications in the Earth Sciences, Geol. Surv. Can.
Paper 89-9.
Hill, M.O., 1979. DECORANA - a FORTRAN program for detrended correspondence analysis and
reciprocal averaging: Ecology and Systematics. Cornell Univ. Ithaca, New York, 52 pp.
Hohn, M.E., 1978. Stratigraphic correlation by principal components: effects of missing data. J. Geol.,
86: 524-532.
Hohn, M.E., 1985. SAS program for quantitative stratigraphic correlation by principal components.
Computers and Geosciences, 11: 471-477.
414
Howell, J.A., 1983. A FORTRAN 77 Program for automatic stratigraphic correlation: Computers and
Geosciences, 9: 311-327.
Hudson, C.B. and Agterberg, F.P., 1982. Paired comparison models in biostratigraphy. J. Math. Geol.
14: 141-159.
Jackson, A., Lew, S.N. and Agterberg, F.P., 1984. DISSPLA program for display of dendrograms from
RASC output. Computers and Geosciences 1 0 59-165.
Jasko, T., 1984. The first find; estimation of the precision of range zone boundaries. Computers and
G~osc.,10: 133-136.
Jeletzky, J.A., 1965. Is it possible to quantify biochronological correlation? J . Paleont., 39: 135-140.
Jenkins, G.M. and Watts, D.G., 1968. Spectral Analysis and its Application. Holden-Day, San
Francisco, 525 pp.
Johnson, N.I. and Kotz, S., 1969. Discrete Distributions. Houghton Mifflin Company, Boston,
Massachusetts, 328 pp.
Jones, D.J., 1958. Displacement of microfossils. J . Sediment. Petrol., 28: 453-467.
Kemp, F., 1982. An algorithm for the stratigkraphic correlation of well logs. J. Math. Geol., 14: 271-285.
Kemple, W.G., Sadler, P.M. and Straws, 1990. A prototype constrained optimization solution to the time
correlation problem. In: F.P. Agterberg and G.F. Bonham-Carter (Editors), Statistical
Applications in the Earth Sciences, Geol. Surv. Can. Paper 89-9.
Kendall, M.G., 1975a. Rank Correlation Methods. Griffin, London, 202 pp.
Kendall, M.G., 197513. Multivariate Analysis. Hafner, New York, N.Y., 210 pp.
Kendall, M.G. and Stuart, A,, 1961. The Advanced Theory of Statistics, Volume 2. Hafner, New York,
676 pp.
Kent, D.V. and Gradstein, F.M., 1985. A Jurassic and Cretaceous geochronology. Geol. SOC.America
Bull., 96: 1419-1427.
Kent, D.V., and Gradstein, F.M., 1986. A Jurassic to Recent Chronology. In: P.R. Vogt and
B.E. Tucholke (Editors), The western Atlantic region, Vol. M, The Geology of North America,
Geol. Soc. Am., pp. 45-50.
King, C., 1983. Cainozoic micropaleontological biostratigraphy of the North Sea. Rept. Inst. Geol.
Sciences No. 82/7,40 pp.
Kwon, B.D. and Rudman, A.J., 1979. Correlation of geologic logs with spectral methods. Math. Geology,
11: 373-390.
Lapin, L.L., 1982. Statistics for Modern Business Decisions, 3rd Edition. Harcourt, Brace, and
Jovanovich, Inc., New York, N.Y., 887 pp.
Lerche, I., 1990. Philosophies and strategies of model building. In: T.A. Cross (Editor), Quantitative
Dynamic Stratigraphy, Prentice Hall, Englewood Cliffs, New Jersey, pp. 21-44.
McKenzie, R.M., 1981. The Hibernia- a classic structure. Oil and Gas J., September, 1981, pp. 243-247.
McKerrow, W.S., Lambert, R.St.J. and Chamberlain, V.E., 1980. The Ordovician, Silurian and
Devonian time scales. Earth Plan. Sc. L., 51: 1-8.
McLaren, D.J., 1978. Dating and correlation, a review. In: G.V. Cohee, and others (Editors),
Contributions to the geologic time scale. American Ass. Petroleum Geologists, Studies in
Geology 6, pp. 1-7.
Macellari, C.E., 1986. Late Campanian-Maastrichtian ammonite fauna from Seymour Island (Antarctic
Peninsula). J . Paleont., 60,": 1-55.
Magara, K., 1976. Thickness of removed sedimentary rocks, paleopressure, and paleotemperature,
southwestern part of western Canada Basin. Am. Assoc. Petroleum Geologists Bull., 60: 554-565.
Maher, L.J.. 1972. Nomograms for computing 0.95 confidence limits of pollen data. Rev. Palaeobotany
Palynology, 23: 85-93.
Maher, L.J., 1981. Statistics for microfossil concentration measurements employing samples spiked with
marker grains. Rev. Palaeobotany Palynology, 32: 153-191.
Mann, C.J. and Dowell, T.P.L., Jr., 1979. Quantitative lithostratigraphic correlation of subsurface
sequences. Computers and Geosciences, 4 295-306.
Menning, M., 1989. A synopsis of numerical time scales, 1917-1986. Episodes, 12(1): 3-5.
Millendorf, S.A., Brower, J.C. and Dyman, T.S., 1978. A comparison of methods for the quantification of
assemblage zones. Computers and Geosciences, 4 229-242.
Miller, F.X., 1977. The graphic correlation method in biostratigraphy. In: E.G. Kauffman and
J.E. Hazel (Editors), Concepts and methods of biostratigraphy, Dowden, Hutchison and Ross, Inc.,
Stroundsburg, USA, pp. 165-186.
Miller, K.G. and Fairbanks, R.G., 1985. Cainozoic 6 1 8 0 record of climate and sealevel. S. Afr. J. Sci., 81:
248-249.
Miller, R.G., 1974. The Jackknife - a review. Biometrika, 61: 1-17.
 415

Mohan, M., 1985. Geohistory analysis of Bombay High region. Marine and Petroleum Geology, 2:
350-360.
Mosteller, F., 1951. Remarks on the method of paired comparisons, I, The least squares solution
assuming equal standard deviations and equal correlations. Psychometrika, 16: 3-9.
Mouterde, R., Ruget, C. and Tintant, H., 1973. Le passage Oxfordien - Kimmeridgien au Portugal
(regions de Torres-Vedras et du Montejunto). Com Ren. Acad. Sc. Paris, 277 (SBr. D): 2645-2648.
Muller, C. and Willems, W., 1981. Nannoplankton en planktonische foraminiferen uit de Ieper-Formatie
(Onder-Eoceen)in Vlaanderen (Belgie). Natuurw. Tijdschr., 62: 64-71.
Nazli, K., 1988. Geostatistical modelling of microfossil abundance data in upper Jurassic shale, Tojeira
sections, central Portugal. Unpublished M.Sc. thesis, Univ. Ottawa, 369 pp.
Nowlan, G.S., 1986. Paleontology: ancient and modern. Geoscience Canada, 13 (2): 67-72.
Odin, G.S. (Editor), 1982. Numerical Dating in Stratigraphy, Parts I and 11. Wiley- Interscience,
Chichester, 1040 pp.
Olea, R.A., 1988. Correlator - an interactive computer system for lithostratigraphic correlation of
wireline logs. Kansas Geol. Survey, Lawrence, Kansas, Petrophysical Ser. 4,85 pp.
Oleynikov, N.A. and Rubel, M. (Editors), 1988. Quantitative Stratigraphy - Retrospective Evaluation
and Future Development. Institute ofGeology, Acad. Sc. Estonian SSR,Tallinn, U.S.S.R., 167 pp.
Palmer, A.R., 1954. The faunas of the Riley formation in central Texas. J . Paleont., 28: 709-786.
Postuma, J.A., 1971. Manual of Planktonic Foraminifera. Elsevier, Amsterdam, 420 pp.
Quenouille, M., 1949. Approximate tests of correlation in time series. J . Royal Statist. Soc. Ser. B., 11:
18-84.
Rao, C.R., 1973. Linear Statistical Inference and its Applications. Wiley, New York, N.Y., 625 p.
Reinsch, C.H., 1967. Smoothing by spline functions. Numerische Mathematik, 10: 177-183.
Reinsch, C.H., 1971. Smoothing by spline functions. 11. Numerische Mathematik, 16: 451-454.
Reyment, R.A., 1980. Morphometrical Methods in Biostratigraphy Academic Press, London, 175 pp.
Reyment, R.and Sturesson, U.,1987. Correlation of chemical and physical environmental fluctuations
in a late Cretaceous borehole sequence - A multivariate study. Sed. Geol. 53: 311-325.
Riedel, W.R., 1979. Recent and potential advances in DSDP biostratigraphy. Am. Ass. Petr. Geol. Bull.,
63: 516.
Roberts, F., 1976. Discrete Mathematical Models. Prentice-Hall, Englewood Cliffs, N.J., 559 p.
Roberts, F., 1978. Graph Theory and its Applications to Problems of Society. Regional Conference Series
in Applied Mathematics 29, SIAM, Philadelphia, Penn., 122 pp.
Royden, L., Sclater, J.G. and Von Herzen, R.P., 1980. Continental margin subsidence and heat flow:
Important parameters in formation of petroleum hydrocarbons. Bull. Am. Assoc. Petr. Geol., 64:
173-187.
Russell, D.A., 1975. Reptilian diversity and the Cretaceous-Tertiary transition in North America. Geol.
Ass. Can. Spec. Paper 13: 119-136.
Russell, D.A., 1977. The biotic crisis a t the end of the Cretaceous period. National Museums of Canada,
Syllogeus, no. 12, pp. 11-23.
Rubel, M., 1978. Principles of construction and use of biostratigraphical scales for correlation.
Computers and Geosciences, 4 243-246.
Rubel, M. and Pak, D.N., 1984. Theory of stratigraphic correlation by means of ordinal scales.
Computers and Geosciences, 10: 97-105.
Salin, Yu. S.,1989. Computerized stratigraphic correlation by means of a geochronological scale. In: A.
Oleynikov and M. Rubel (Editors), Quantitative Stratigraphy-Retrospective Evaluation and
Future Development, Acad. Sciences Estonian S.S.R., Institute of Geology, Tallinn, pp. 73-80.
Sankoff, D. and Kruskal, J.B. (Editors), 1983. Time Warps, String Edits, and Macromolecules: The
Theory and Practice of Sequence Comparison. Addison Wesley, London, 382 p.
Schindewolf, O.H., 1950. Grundlagen und Methoden der palaontologischen Chronologie, 3rd Ed.
Borntraeger, Berlin, 152 pp.
Schlumberger, 1979. Log Interpretation Charts. Schlumberger Ltd., New York, 92 p.
Schoenberg, I.J., 1964. Spline functions and the problem of graduation. Proc. National Academy of
Sciences of the U S A . , 52: 947-950.
Schwarzacher, W., 1985a. Principles of quantitative lithostratigraphy - the treatment of single sections.
In: Quantitative Stratigraphy, UNESCO, Paris and Reidel, Dordrecht, pp. 361-386.
Schwarzacher, W., 1985b. Lithostratigraphic correlation a n d sedimentation models. In:
F.M. Gradstein et al., Quantitative Stratigraphy, UNESCO, Paris, and Reidel, Dordrecht,
pp. 387-418.
Sclater, J.C., and Christie, P.A.F., 1980. Continental stretching: a n explanation of t h e post
mid-Cretaceous subsidence of the central North Sea basin. J. Geophys. Res., 85: 371-379.
416
Shaw, A.B., 1964.Time in Stratigraphy. McGraw-Hill, New York, 365 pp.
Shaw, B.R., 1978. Parametric interpolation of digitized log segments. Computers and Geosciences, 4:
277-283.
Signor, P.W. and Lipps, J.H., 1982. Sampling bias, gradual extinction patterns and catastrophes in the
fossil record. In: L.T. Silver and P.H. Schulz (Editors), Geological Implications of Impacts of Large
Asteroids and Comets on the Earth. Geol. SOC. Am., Special Pap. 190,pp. 291-296.
Silverman, B.W., 1984. A fast and efficient cross-validation method for smoothing parameter choice in
spline regression. J. American Statistical Ass., 79:584-589.
Smith, D.G. and Fewtrell, M.D., 1979. A use of network diagrams in depicting stratigraphic time
correlation. Geol. Soc. London J., 136: 21-28.
Smith, T.F. and Waterman, M.S., 1980 New stratigraphic correlation techniques. J . Geol. 88: 451-457.
Southam, J.R., Hay, W.W. and Worsley, T.R., 1975. Quantitative formulation of reliability in
stratigraphic correlation. Science, 188: 357-359.
Springer, M. and Lilje, A,, 1988. Biostratigraphy and gap analysis: the expected sequence of
biostratigraphic events. J . Geol., 96: 228-236.
Srivastava, S.P. (Editor), 1986. Geophysical maps and geological sections of the Labrador Sea. Geol.
Survey Canada, Paper 85-16,llpp.
Stainforth, R.M., Lamb, J.L., Luterbacher, H., Beard, J.H. and Jeffords, R.M., 1975. Cenozoic planktonic
foraminifera zonation and characteristics of index forms. Univ. Kansas Paleont. Contr., no. 62,
pp. 1-162.
Stam, B., Gradstein, F.M., Lloyd, P. and Gillis, D., 1987. Algorithms for porosity and subsidence history.
Computers and Geosciences, 13 (2).
Stam, B., 1987. Quantitative Analysis of Middle and Late Jurassic Foraminifera from Portugal and its
Implications for the Grand Banks of Newfoundland. Utrecht Micropaleontological Bull. 34,
167 pp.
Strauss, D. and Sadler, P.M., 1989. Classical confidence intervals and Bayesian probability estimation
for ends and local taxon ranges. Math. Geol., 21: 411-427.
Sullivan, F.R., 1965. Lower Tertiary nannoplankton from the California Coast Ranges; 11. Eocene.
Univ. Calif. Publ. Geol. Sc.,53: 1-52.
Thomas, F.C., Gradstein, F.M. and Griffths, C.M., 1988. Bibliography and Index of Quantitative
Biostratigraphy. Special Publ. No. 1, Comm. Quantitative Stratigraphy, Bedford Inst. Oceanogr.,
Dartmouth, N.S., Canada, 58 pp.
Tipper, J.C., 1988. Techniques for quantitative stratigraphic correlation: a review and annotated
bibliography. Geol. Mag., 125 (5):475-494.
Tjalsma, R.C. and Lohmann, G.P., 1983. Paleocene - Eocene bathyal and abyssal benthic Foraminifera
from the Atlantic Ocean. Micropal., Spec. Publ., no. 4,76pp.
Tocher, K.D., 1950. Extension of the Newman-Pearson theory of tests to discontinuous variates.
Biometrika, 37: 130.
Tukey, J., 1958. Bias andconfidence in not quite large samples. Annals Math. Statist., 29: 614.
Tukey, J.W., 1977. Exploratory Data Analysis. Addison-Wesley, Reading, Massachusetts, 688 pp.
Utreras, F., 1981. Optimal smoothing of noisy data using spline functions. SIAM J. Stat. Comput., 2:
349-362.
Vail, P.R. and Mitchum, R.M., Jr., 1979. Global cycles of relative changes of sea-level from seismic
stratigraphy. Am. Ass. Petr. Geol. Mem. 29: 469-472.
Vail, P.R., Mitchum, R.M., Jr. and Thompson, S.,111, 1977. Seismic stratigraphy and global changes of
sealevel. Part 4. Mem. Am. Assoc. Pet. Geol. 26: 83-97.
Van Hinte, J.E., 1978. Geohistory analysis, application of micropaleontology in exploration geology.
Am. Assoc. Petrol. Geol. Bull., 62: 201-227.
Van Hinte, J.E., 1984. Synthetic seismic sections from biostratigraphy. Am. Ass. Petr. Geol. Mem. 34:
674-685.
Van Valen, L. and Sloan, R.E., 1977. Ecology and the extinction of the dinosaurs. Evolutionary Theory,
2: 37-64.
Vrbik, J., 1985. Statistical properties of the number of runs of matches between two random
stratigraphic sections: Mathematical Geology, 17: 29-40.
Watts, A.B., and Steckler, M.S., 1981. Subsidence and tectonics of Atlantic-type continental margins.
Oceanologica Acta, vol. 4,suppl. 1981,no. SP, pp. 143-153.
Wahba, G., 1975. Smoothing noisy data with spline functions. Numerische Mathematik, 2 4 383-393.
Waterman, M.S. and Raymond, R., Jr., 1987. The match game: new stratigraphic correlation
algorithms. Math. Geol. 19: 109-127.
 417

Waterman, M.S., Smith, T.F. and Beyer, W.A., 1976. Some biological sequence metrics. Adv. Math., 2 0
367-387.
Wegman, E.J. and Wright, I.W., 1983. Splines in statistics. J. American Statistical Ass., 78: 351-365.
White, J.M., 1990. Exploration of a practical technique to estimate the relative abundance of rare
palynomorphs using an exotic spike. In: F.P. Agterberg and G.F. Bonham-Carter (Editors),
Statistical Applications in the Earth Sciences, Geol. Surv. Can. Paper 89-9.
Whittaker, E.T., 1923. On a new method of graduation. Proc. Edinburg Math. SOC.,41: 63-75.
Wilkinson, E.M., 1974.Techniques of data analysis - seriation theory: Archaeo- Physika, 5: 1-142.
Williams, D.F., 1990. Selected approaches of chemical stratigraphy to time-scale resolution and
quantitative dynamic stratigraphy. In: T. A. Cross (Editor), Quantitative Dynamic Stratigraphy,
Prentice Hall, Englewood Cliffs, New Jersey, pp. 543-565.
Williams, D.F., Lerche, I . and Full, W.E., 1988. Isotope chronostratigraphy: Theory and Methods.
Academic Press, San Diego, 352 pp.
Williamson, M.J., 1987. Quantitative biozonation of the Late Jurassic and Early Cretaceous of the East
Newfoundland Basin. Micropaleontology, 33: 37-65.
Williamson, M.A. and Agterberg, F.P., 1990. A quantitative foraminifera1 correlation of the late
Jurassic and early Cretaceous offshore Newfoundland. In: F.P. Agterberg and C.F. Bonham-
Carter (Editors), Statistical Applications in the Earth Sciences, Geol. Surv. Can. Paper 89-9.
Wilson, L.R., 1964. Recycling, stratigraphic leakage and faulty techniques in palynology. Crana
Palynologica, 5: 427-436.
Wold, S.,1974. Spline functions indata analysis. Technometrics 16 (1):1-11.
Wood, R.I., 1981. The subsidence history of the Conoco well 15/30-1,Central North Sea, Earth and
Planetary Sci. Lett., 54: 306-312.
Worsley, T.R. and Jorgens, M.L., 1977. Automated biostratigraphy. In: A.T.S. Ramsay (Editor),
Oceanic Micropaleontology, Academic Press, London, 2:1201-1229.
Ziegler, P.A., 1981. Evolution of Sedimentary basins in Northwest Europe. In: L.V. Illing and
G.D. Hobson (Editors), Petroleum Geology of the Continental Shelf of Northwest Europe. Inst. of
Petroleum, London, pp. 3-39.
This Page Intentionally Left Blank
 419

INDEX Brunk, H.D., 175

Burial history, 11,364
Burroughs, W.A., 75,76,405
BURSUB computer program, 364
Buzas, M.A., 6
Byron, D.N., 389,407
Adjacency matrix, 62 C language, 406
Adolphus D-50 well, 125, 284, 286, 315-320, 335- Cambrian, 3
347 Canadian Atlantic Margin, 118-132, 351-366, 382-
Age determinations, 72,98-102 387
-depth diagram, see event-depth diagram Carinati, R., 47
Alveolinids, 268-275 Carr, P.F., 102
Ammonite zones, 52,96-98 C a d , B., 61
Anomalous events, 260,307 CASC applications, 351-371,382-387
Arcs, 62 -computer program, 12,320-338,389-407
ARIMA method, 55,56 -1: Event-depth curves module, 394,402,403
Armstrong, R.L., 98 -2: Statistical analysis module, 395,403
Arrow of time, 19,41 -3: Multi-well comparison module, 395,396,404
Ascoli, P., 126
Assemblage zone, 3,5,7,20 Cenozoic calcareous plankton datum events, 125-
129,354-356
Aubry, M.-P., 77 -Foraminifera, 6
Autocorrelation, 5458,260-268 - foraminifera1dictionary, 121-122
Automated correlation, 107,311-387 --zonation, 126,374
Average interval zone, 22 -optimum sequence, 185,318,374
Axiomatic approach, 27
- time-scale, 318,374
Barrell, J., 16 Central limit theorem, 43
Bartlett’s chi-squared test, 290,291 Central North Sea, 371-381
BASIC, 390 Chamberlain, V.E., 102
Basin analysis, 1, 311 Cheetham, A.H., 20
Bassett, M.G., 15 Chemical events, 17
Baumgartner] P.O.,7,39,261,310 Chi-squared test, 263,293,338
Beard, J.H., 126 Christopher, R.A., 29
Beaver. R.J., 60.141.224,250,276 Christie, P.A.F., 372
. . Chronogram, 78-85,92
Beckinsale, R.D:, 102 Chronostratigraphic correlation, 13
Benthonic Foraminifera, 6,10,55,127,371-381 Chronostratigraphy, 1,5,11-13,25
Bentonites, 132
Berge, C., 6 1 Chronozones, 77
Berger, W.H., 45 Chung, C.F., 405
Berggren, W.A., 1,77,126,127,354,355,371-373,Clark, R.M., 15
Clique, 63, 117
376,378,380
Bernoulli trial, 47,48 Cluster analysis, 74
Best fit channel, 314 Clustering in time, 10,23,184
Beyer, W.A., 15 Coding, 103-139
Binomial models, 49-59,153,223-227 Coeval events, 37,152,175-178
-test for randomness, 9,48-49,143-145 Composite standard method, 2,8,312-314
Biochronology, 320,381 Computer programs. 8.389-407
Biostratigraphic assemblage zones, 3,5,7,20 - simulation experiments, 85-92, 204-214, 339,
-correlation, 25 347-350
-event, 29-30 -terminal, 9,14
-resolution. 312 Computers & Geosciences, 404,405
- zonations,.l, 13 Concurrent range zone, 5,21
Biostratigraphy, 1,5-11 ---,multi-taxon, 21,22
Blank,R.G.,48,145,163-165,266,311 Confidence limits, 350
Bliss, C.I., 194 -
CONISS computer program, 74
Blow, W.H., 126 Conservative ranking methods, 165-169
Bonham-Carter, G.F., 4,75,76,373 -zonations, 6,294
Box, G.E.P., 55 Constrained seriation, 76
Bramlette, M.N., 108 Co-occurrences, 28
Co-processor, Math, 389-390,406
Briggs, S.,405 Correlation and Scaling in time (CASC), see CASC
Brinkmann, R.,8 Correspondence analysis, 74,75,373
Brower, J.C., 1, 4, 20, 74-76, 120, 141, 179, 184, Cowie,J.W., 15
185,206,239,260,274,298-300,311,317,354,
Cox, A.V.,16.77-84,86.92-102
405 Craven, P., 340
420

Cretaceous microfossils, 358-366 Epibole, 42

- nannofossils, 6 Error analysis, 10,98-102
-Tertiary boundary, 36,37 -bar, 12,327
Cross, T.A., 17 Eubank, R.L.,68,311,338,349
Cross-association, 14,68 Event, 20,390
Cross-over frequency, 23,60,183,191 -depth diagram, 12,311-388
Cross-plots, 367,395 -level, 105,316
Cross-validation method, 14,339-342 Evolutionary sequence, 5,73
Cubic polynomial (spline) curves, 67,92-98 EXE files, 389
- spline fitting, 9,286,323,347-350,395 Exit of taxon, 8,20
-spline function, 67,94 Exploration micropaleontology, 28
Cubitt, J.C., 4 -wells, 119,358,359,372,381
Culver, S.J., 6 Explorationists, 12
Curve-fitting, 67 Exponential autocorrelation model, 56,57
Cycles, 170-175,209,392
F-matrix, 146
Dalrymple, G.B., 77-79,84 Facies, 26
DATfile, 103,105-106,112-118,123,391 Factor analysis, 74
Data input module, 391,396,397 Fairbanks, R.G., 16,17
Databases, 264 Faunal method, 26
Davaud, E., 7, 30-32,61, 65, 108, 109, 116-1.18, Fearon, M., 8,364,407
141,165,175,178,259,268-275,299 Ferrer, J., 413
David, H.A., 60,141,227 Fewtrell. M.D.. 61
David, M., 59 File management, 103-139
Davidson, R.R., 60 Filter, 67
Davidson Model, 60 Final reordering, 199
Davis, J.C., 14,43 Finney, D.J., 194
Deboo, P.B., 20 First appearance datum (FAD), 8,273
De Boor, C., 9,68-70,311,338,348,395,407 -consistent occurrence, 116
DECORANA computer program, 75 Fisher, R.A., 194
Deep Sea Drilling Program, 3,25 fly^, J.J., 77,354,380
Defaults, 142,396 Foraminifera, 6,53,54
Dendrogram, 23 Forbidden structure, 65
Dennison, J.M., 50,51,58 FORTRAN programs, 320,338,348,389-407
DEN0 computer program, 184 Fossil taxa, 20
Depth file (DEP file), 103,106-107,391 Foster. N.H.. 27
Deterministic amroach. 165 Fractile of normal distribution in standard form,
Dictionary file (DIC file), 103,104-105,391 60,87,139.189.228
Dienes, I., 47 Frequency distributions of stratigraphic events,
Dinoflagellates, 6,382-387 37-45,287-295
DIorio, M.A., 10,75, 76, 181, 198,267, 280, 296, -ofoccurrence, 6,7,129-132
381-387,390,406,407 Fulkerson, D.F., 64
Dirichlet distribution, 33 Full, W.E., 16,17
DISSPLA language, 184
Distance (D) method, 183,189,393 Gale, N.H., 102
-option (mainframe CASC), 321 Gap problem, 15
Directed graph, 62 Gaussian distribution, 39,43,132,187
Dixon, W.J., 132,133,135 Generated subgraph, 64
Doeven, P.H., 116,405 Generalized cross-validation, 340
Doveton, J.H., 11 Geochronologic resolution, 76,85
Dowell,T.P.L., Jr., 14 Geochronology, 11,16,26
Drill cutting samples, 28,38,127 Geohistory analysis, 11
Drobne, K., 259,268-275 Geological correlation, 24
Drobne’s alveolinids, 268-275,295-305 - time-scale, 16,92-98
Drooger, C.W., 9,25,311 GEOSCI computer programs, 406
Dummy number, 120 Geostatistics, 59
Duris, C.S., 395 Gill, D., 4
Dyman, T.S., 74 Gillis, D., 11, 389,407
Gilmore, P.C., 65
Edge (in maDh theory). 62 Glauconite dates, 98-101
Ediards,L.E., 8,38:40,45,60,141,165,204,224, Glenn, W.A., 60,227
250,276,305-310,405 Glenn-David model, 60,227-238
Efron, B.,343 Global error bar, 12,327
Ellis. C.F.. 48.163-165 -sea level changes, 16,366,380
Entry of taxon, 8,20 Colub, G.H., 340
Eocene, 75,112-115,354,355,377 Gordon, A.D., 14,15
 42 1

Gradstein, F.M., 1,2,4,7-12, 16,49,52,54,55, 58, Interpolation spline, 67,70

68, 71, 75-77, 96-101, 116, 118-132, 179, 181, Interval graph, 64,117
184,185,206, 226,227,260,288, 299,311-320, -zone, 21
354-356,371-381,389,404-407 Isochron contouring, 1,12
Gradstein, W.S., 260,389,405,406 Isotope chronostratigraphy, 16-17
Gradstein-Thomas database, 118-132, 281, 284- IUGS Commission on Stratigraphy, 1,15
295,309
Grand Banks, 13,75,118-132,382-387
Graph theory, 2,7,61-66,170 Jackknife method, 251-258,342-347
Griffiths, C.M., 4 -scaling module, 393,400,401
Grimm, E.C., 74 Jackson, A,, 184,405,406
Gross, O.A., 64 Jasko,T., 31,33,35,36,38
Guex, J.,2,7,61,62,65,66,108,109,113,116-118, Jeffords, R.M., 126
141,165,175,178,259,268-275,276,299 Jeletzky, J.A., 26
Guex levels, 118 Jenkins, G.M., 55,57
Gyji, R.A., 98 Jenkins, W.A.M., 126
Johnson, N.I., 58,224,225
Hald, A,, 225,290,291,338 Johnston, A., 405
Hallam, A,, 95,96 Jones, B.G., 102
Haq, D.U., 99,100,380 Jones, D.J.,27
Hardenbol, J., 99,100,380 Jorgens, M.L., 145,170,171,174
Harland, W.B., 16,77-83,86,92-102 Jurassic, 3,52,72
Harper, C.W., Jr., 26-28, 156, 166, 168, 204, 239, -Cretaceous boundary, 16,77,98
246-250 - radiolarians, 261
Hat matrix, 349
Hay, W.W., 9,48,50,51,58,108-115,141-152,225, Kaminski, M.A., 1,371,374,376,378
305-309,311,313 Kemp, F., 14
Hay example, 49, 108-118, 141-163, 191-201, 223, Kemple, W.G., 163
257 Kendall, M.G., 74,80,82,141,163,177,239-2 46
Hazel, J.E., 74 Kendall's rank correlation coefficient (tau), 239-
Heath, G.R., 45 250,277
Heath, M., 340 Kent, D.V.,16,77,96-101,354,380
Hedberg, H.D., 4,20,21,22 King, C., 376
Heller, M., 260,389,405,406 Kirk, J., 407
Hemelrijk, J., 175 Klitgord, K.D., 77
Hemera, 42 Knox, R.B., 373-378
Hewitt, R., 77 Koch, C.F., 6
Hiatus, stratigraphic, 1,184,289 Kotz, S., 58,224,225
Hibernia Oilfield, 13,358-366 Kriging, 67
Hibbert. P.. 395.407 Kruskal, J.B., 15
Highesto&urrence, 1,31-45,294,300 Kwon, B.D., 14
Hill, M.O., 75
Hoffman, A.J.. 65 Labrador Shelf, 12,75,118-132,382-387
Hohn, M.'E., 74 Lamb, J.L., 126
Howell, J.A., 15 Lambert, R.St. J., 102
Hudson, C.B., 60,276,405 Lapin, L.L., 43
Huang, Z., 364,389,407 Last Appearance Datum (LAD), 8,273
Hydrocarbons, 11 -consistent occurrence, 116
Least squares, method of, 8,93,280
IGCP Project No. 148,1,2-5,76,405 Lerche, I., 1,11,16,17
IGCP Catalogue, 4 Lessard. R.. 405
IMSL Library, 320 Lew, S.N., ~Z,~68,118,184,311,320,389,406,
407
Index fossil, 5,26,221 Lilje. A., 31
Indian Harbour M-52 well, 330-335 Line ofcorrelation, 8,313,317
Indirect distance estimates, 2,180,190 -of observation, 316
-method of spline-fitting, 68 Linnean name, 20
Initial Unitary Associations (IUA), 66,268 Lipps, J.H., 35,36
Integration of datasets, 10,382-387 Lithosome, 14
Interactive CASC session Lithostratigraphic correlation, 14
-computer program, 14,405 Lithostratigraphy, 1,5,1415,19
Inter-event distance, 192 Llewellyn, P.G., 16,77-83,86,92-102
Inter-fossil distance, 184 Lloyd, P., 11
International Geological Correlation Programme Local error bar, 12,325
(IGCP) Project 148,1,2-5,76,405 -range zones, 30,31
International Union of Geological Sciences (IUCS) Log-likelihood function, 80,82,84-85,88
Stratigraphic Guide, 4 Lohmann, G.P., 354
422
Lowest occurrence, 31-45,300 Odin, G.S., 16,77,98,99
Luterbacher, H., 126 Ogg, J.G., 77
Olea, R.A., 14
McDowell, F.W., 98 Oleynikov, N.A., 4
McKenzie, R.M., 364 Oligocene, 355,377,378
McKerrow, W.S., 102 Oliver, E.M., 405
McLaren, D.J., 24 Oliver, J., 12,68,311,320,406
Macellari, C.E., 32,34 Oppel,A., 116
Magara, K., 414 Oppel zone, 7,21,22,116,268
Magnetostratigraphy, 15 Optimum clustering, 23
Maher, L.J., 51,52 -spline curve, 338
Mann, C.J., 14,47 -sequence, 10,141,143,157
Marini, A., 47 Ordering, see Ranking
Marker horizon, 132,219-221,298 Oxfordian, 49,98
Mass extinction, 35,36,289
Massey, F.J., 132,133,135 P/B ratio, 73
Matching, 14 P-matrix., 149.193
, -
Mathematical statistics, 47-102 Paired comparison models, 226
Matrix (matrices), 146, 147 Pak, D.N., 165
Maximal clique, 63 Paleocene, 112-115,354,377
-horizon, 117 Paleontological record, 20
Maximum likelihood, method of, 56,80,89,99 Paleo-water depth, 11
Menning, M., 16,102 Palmer. A.R.. 259.276.305.366-371
~.
~ ~

Merriam, D.F., 4 Palmer's database, 275-280,305-310,366-371

Mesozoic Foraminifera, 358-366 PARfile, 103,106,391,396
- time-scale, 92-98 Peak occurrence, 8
Microcomputers, 9,389,406 Penalty points, 243
Microfossil abundance data, 49-59,67-73 Phanerozoic time-scale, 77
Micro-RASC, 321,391-396 Phylozone, 21
Microsoft Disk Operating System (DOS), 103 Pickton, C.A.G., 16,77-83,86,92-102
Millendorf, S.A., 74 Planktonic foraminifers, 127,354,374
Miller, F.X., 314 Plotting language DISSPLA, 184
Miller, K.G., 16,17,77 Poisson distribution, 33,50
Miller, R.G., 258,343 - -, compound, 33
Miocene, 75,355,378 Polynomial, 67
Missing data, 159,160 Population, biological, 20
Mitchum, R.M., Jr., 16,365,366 -statistical, 29,207
Mixing of sediments, 7,27,44-45 Postuma, J.A., 126
Modified Hay method, 172,175 Potenza, R.G., 47
-local error bar, 12,327 Preprocessing module, 391,397,398
- RASC method, 280-310,407 Presorting, see Probabilistic ranking
--module, 393,394,401 Price, R.J., 405
Mohan, M., 12 Principal component analysis, 74
Moore, P., 405 Probabilistic biostratigraphy, 9,108
Morton, A.C., 373-378 -ranking, 156-161,246-250,392
Mosteller, F., 60 Probit, 71
Mouterde, R., 52 Prothero, D.R., 77
Muller, C., 354 Pseudo-random normal number generator, 135,
Multiple pairwise comparison, 9,60-61,226-227 204
Multivariate analysis, 4,73-76 Pseudovalues, 343
Multi-well comoarison., 12.31
, 1-387
Q-mode, 54,73
Nannofossils, 108-112 Quantitative dynamic stratigraphy, 17
Nazli, K., 49,52,55-58,70,71 -stratigraphic correlation, 1
Nel, L.D., 9, 37, 61, 145, 156, 159, 179, 228, 229, -stratigraphy, 2,19-45
247,249,260,389,404,405 --, Committee on, 4,389
Noise, statistical, 6,11,26,57,59,67-73 Quenouille, M., 342
Normal (Gaussian) probability distribution, 43, Quinn, B.G., 102
79,187,293
Normality test, 10,215-219,259-280 R-Matrix, 149
-- module, 393,400 R-Mode, 54,74
Nowlan, G.S., 19 Radiolarians, 6.8
Numerical time-scale, 16,76-102 Radiometric ages, 15
Random variability, 59
Obradovitch, J., 77 -normal numbers, 132-139,201-214
Occurrence table, 215,392 -number generator, 135,247
 423

Range chart, 24,30-31, 116,294,309 Shaw, B.R., 14

-through method, 20,116 Shih, K.C., 405
-zone, 21 Signal-plus-noise model, 57
Rank correlation, 239-250 Signor, P.W., 35,36
-evaluation module, 392,393,400 Silurian, 3,102
Ranked sequence, 141 Silverman, B.W., 340
Ranking, 141-178,246-250 Skewness, 38,42,290
-and Scaling (RASC), see RASC Sloan, R.E.,37
-module, 392,398,399 Slotting method, 14
Rao, C.R., 86,91 Smith, A.G., 16,77-83,86,92-102
Rao, S.N.,4 Smith, D.G., 61
RASC biochronology, 185,371-381 Smith, T.F., 15
- biozones, 23 Smoothing factor (SF),68,322,347
-computer program, 75,76, 157,389-407 -spline, 67-73
-distance, 186-198 Sohl, N.F., 6
-method, 10 Southam, J.R., 48,141,225
-normality test, 215-219.301 Spearman’s rank correlation coefficient (rho), 239-
-scale, 226-227 243
Regional time-scale module, 394,401 SPLIN computer program, 390,407
Reworking, 27,128,289,292 Spline-curve fitting
Reyment, R., 4, 15,73,74 Spores and pollen, 74,382-387
Reinsch, C.H., 68, 311, 338 Springer, M., 31
Reinsch-De Boor spline-fitting, 347 Srivastava, S.P., 126,354, 387
Reinsch’s suggestion, 338 Stainforth, R.M., 126
Riedel, W.R., 312 Stage boundaries, age of, 77,85-92
Riley Composite Standard (RST),367 Stam, B., 11,49,52-55,70-73
Roberts, F., 61,63,64, 118 Statistical model, 186-201
Royden, L., 415 Steckler, M.S., 416
Rubel, M.,4,141, 156, 165-169, 175, 178 Steno, law of, 19
Rudman, A.J., 14 Step model (RASC), 209,215,242-246,392
Ruget, C., 52 Stratigraphic concepts, 1-45
Russell, D.A., 37 -correlation, 1,24
-leaks, 27
S-Matrix, 149 -relationships, 20
s-ratio, 65,66,268 Stratigraphy, 19
Sadler, P.M., 31-34,38,163 Strauss, D., 31-34,38, 163
SAS (Statistical Analysis System), 55,56 Stretching, 4, 14.76
Salin, Yu. S., 175 Strong component, 65
Sampling, 5,6,54 Student’s t-test, 248,254
Sankoff, D., 15 Stuart, A., 80,82
Scaled optimum sequence, 10,185,199-201 Sturesson, U., 74
---,precision of, 250-258 Subjective age-depth data, 358,362
Scaling, 161-165,179-237 -zonation, 128
- module, 392,399,400 Subsidence models, 11
Scattergram, 54,215,392 Sullivan, F.R., 108-118
Schindewolf, O.H., 26 Sullivan database, 107,109,111
Schlumberger, 415 Superpositional relations, 9,27,28, 149-151
Schoenbera, I.J.. 68.338 Switches, 396
SchwarzaAer, W., 1,4,20, 120, 179, 184,185,206,
260,299,311,317,354 T-Matrix, 149
Sclater. J.C.. 372 Taxonomy
Scoring method, 91 Tectronix Advanced Graphics Library, 320
Scotian Shelf, 75,118-132 Tertiary, 108-112,371-381
Second order difference, 215,262 Thomas, F.C., 4,118-132
Sediment accumulation rate, 32,320,325-327,383 Thompson, S., 111, 16
Sedimentation rate (RASC, CASC), see Sediment Thomson, R.,412
accumulation rate Threshold parameter, 6,131,179
Seismic events, 132 Thurstone-Mosteller Model, 60,227
Seismostratigraphy, 16,19,365,366,380 Tie-points, 97, 316
SEQ file, 103,106,124-125,139,300,391 Ties, 149,177
Sequencing methods, 1 Time-scales, 76-101
SER computer program, 298 -series analysis, 55
Seriation, 75,76,298,299 Tintant, H., 52
Set theoretical approach, 4,47 Tipper, J.C., 4, 20,47
Shaw, A.B., 2, 8, 9, 42, 58, 141, 165, 259, 276-280, Tithonian, 98
305-309,311,313-314,366-371 Tjalsma, R.C., 354
424

Tocher, K.D., 175 Watts, A.B., 416

Tojeira sections, Portugal, 53-59,70-73 Watts, D.G., 57
Traceability, 6 , 9 Wegman, E.J., 338
Transitively orientable graph, 65 Weight, 194
Trimming, 280 Weighted distance analvsis. 61. 192-198
Trinomial models, 60-61,223-238 Weighting function, 78,”80,85
Truncated normal distribution, 263 Well logs, 15
TSREG computer program, 390 Westerman, G.E.G.,77
Tukey, J.W., 280,343 White, J.M., 52
White noise, 57
Uncertainty range, 152-154 Whittaker, E.T., 68
Unconformity, 366 Wilkinson, E.M., 75
Undirected graph, 62 Willems, W., 354
Unique event (UE), 221-223,374,391 Williams, D.F., 16, 17
Unitary Associations (UA) method, 2,7, 39, 61-66. Williams. G.L.. 126. 128
116,117,268-275,298,299 Williamson, M.A., 12, 13, 68, 311, 320, 351, 358-
Utreras, F., 340 366,406
Wilson, L.R., 27
Vail, P.R., 16.99, 100,365,366,380 Wold, S., 339,343
Van Couvering, J.A., 354,380 Wood, R.I., 272
Van Hinte, J.E., 11, 12,365,366 Worsley,T.R., 48,145,170,171,174,225
Van Valen, L., 37 Wright, A.J ,338
Variance, analysis of, 292,296 Wright, 1.W ,102
Von Herzen, R.P., 415
Vertex, 62 Yates, F., 194
Virtual co-occurrence, 66
Vrbik, J . , 14 Z-matrix, 193
Z-value, 179-183
Wadge, A.J., 102 Zq-structure, 64
Wahba, G., 338,339,340,349 Ziegler, P.A., 373
Walters, R., 16,77-83,86,92-102 Zonation, planktonic, 354,374
Waterman, M.S., 15 Zone, 20-26